Content SEO

A page is worth creating when it contains something a language model could not have guessed. Everything else in this lesson is downstream of that sentence, and Google has published the vocabulary for the failure it describes.

The mechanism is worth understanding rather than accepting. A model trained on the web has read the top ten results for your query, and several thousand pages very like them. It can produce the average of that set on request, fluently, in seconds, at zero marginal cost. So the average of the set is now free, and anything a reader could have obtained for free is not a reason to visit a website. That is not a moral argument about AI. It is an observation about what remains scarce.

What remains scarce is anything that required somebody to do something. A number you measured. A tool you bought and broke. A decision you made and regretted. A conversation nobody else was in. Google's own phrasing for the standard is unusually blunt: do not recycle what others have already said, or what "could easily be produced by a generative AI model".

Here is the ten-item version of that judgement, and I would rather you got three of them wrong now than on a page you spend a week writing. Three have a surface signal pointing the wrong way, and in all three the tell is the same: effort spent writing is not the same thing as effort spent finding out.

The pattern in the ten is that every non-commodity item required somebody to go and do something before the writing started, and not one of them required more words. The shortest item on that list, three hundred words about a Cloudflare rule, is the only page on the internet with that particular header dump on it. The longest, a four-thousand-word technical SEO guide, is one of several thousand.

One clarification before the rest of the page, because I am about to spend six chapters on gathering evidence and I do not want it read as an argument against writing well. Clarity, structure and sequence are real and they matter, and they are decisions three and five here. What they cannot do is rescue a page with nothing in it. A beautifully organised restatement of the current top three is still a restatement, and it will be indexed, crawled, technically flawless and worth nothing.

Content SEO decides what information should exist and produces it. On-page SEO makes one page communicate its relevance clearly. They are two different jobs and merging them deletes the first one, because only the second is checkable by a plugin.

The reason to separate them is diagnosis rather than tidiness. "Our content is fine, we just need better titles" and "our titles are fine, we just need better content" are two genuinely different diagnoses with two different budgets, and a team that cannot say which column a decision belongs in will spend the year on the cheap one.

Now the part I did not expect to be able to measure. Ahrefs' course splits SEO Content and On-Page SEO into two chapters, which is where the idea for this page came from, so I went and counted whether anybody else does. Eight guides, the seven the brief for this page named plus Google's own starter guide, fetched as raw HTML on the day this went live and sorted heading by heading.

65 of 147 sections are actually about content, which is 44%. Yoast's is the sharpest single row and it is honest about itself: its three named pillars are keyword research, site structure and copywriting, and seven of its thirty-two sections are about content. Two thirds of the guide with the strongest title claim in the set is steps three and five of this roadmap.

The result I expected and did not get belongs here too, because a dataset that only ever agrees with the author is not a dataset. I assumed most of the eight would be dominated by keyword research and on-page, and only 3 of them are. Ahrefs and Semrush's blog post are both comfortably majority-content. The argument for this page rests on the average and on the worst three, not on a claim that everybody is doing it wrong.

I am not accusing anybody of padding. The drift is structural. Keyword placement can be scored, headings can be counted, and "does this page know anything" cannot, so on any shared word count the scoreable half wins. Which is also why this lesson can afford to be about content: the other buckets already have their own pages on this site and all of them shipped before this one.

One more finding from those two results pages, and it decides the URL of the page you are reading. Most guides settle their own slug in a commit message nobody sees. This one is settled here, with the numbers that argue against the choice printed first.

"seo content" carries 3,600 US searches a month and "content seo" carries 1,800. Twice the volume, and not one of the nine pages ranking for the bigger phrasing has it as its own top keyword, while three pages have the smaller one. The market's own pages were built for the phrasing with half the demand, which is the same finding step five produced on a completely different subject. Mature informational markets seem to get collected rather than targeted.

The uncomfortable half is the AI Overview. Both queries return an identical 9-source citation set, and Ahrefs, Yoast and Backlinko are absent from it while a university web team page and a YouTube video are in it. Whatever selects those nine is not selecting on domain authority, and I cannot tell you what it is selecting on. Anybody who can should be asked for their measurement.

Content belongs between deciding where a page lives and deciding how it reads, because it is the last step that can still say no at a reasonable price.

Run the chain out. Keyword research gives you what people want. Search intent gives you what would satisfy them, as a page type with a count behind it. Architecture gives the page a parent and a route in. Content asks what is actually on it, and on-page then makes that legible. Every step narrows, and the cost of reversing goes up at every stage: killing a row in a spreadsheet is free, killing a draft costs a morning, and killing a published page costs a redirect and a small amount of trust.

The usual order is keyword research, then write, then optimize. That order has one structural flaw and it is the one this whole page is about: nobody ever asks whether the page should exist, because by the time anybody is looking at it, it exists. A brief that arrives at a writer has already answered the question by arriving.

And a caveat I would rather give you than have you discover. If your pages have no parent and nothing links to them, content is not your bottleneck and this lesson will not help you. Step five has the audit for that, and it found ninety-seven pages on this site with no editorial link pointing at them. A brilliant page nothing points at is a brilliant page nobody reads.

Every page needs a job it does for the reader and a job it does for the business, and both have to be nameable in a sentence before the page is worth building.

With only the reader job, you get a hobby: useful, well made, and unconnected to anything that pays for it. With only the business job, you get a brochure: a page that exists to be found and does nothing for the person who found it. Both fail, and they fail differently enough that teams argue about which one they have.

The reader job has to be a verb. Not "learn about content decay", which is a topic wearing a job's clothes. "Work out why one of their pages is falling, before spending a day on the wrong fix." That version is checkable: you can read the page afterwards and ask whether somebody could now do it.

The business job has to be something other than traffic. Traffic is not a job, it is a measurement of one. Routes to a service page, proves a capability to somebody deciding whether to hire you, collects an email from exactly the person you want, gets cited by people who influence buyers, or supports a commercial page by being the thing that earns the link. If you cannot name which, the page has no business job and you will not miss it when it is gone.

reader job    "work out why one of their pages is falling"     ← a verb, checkable
business job  "routes to the refresh service, proves I diagnose" ← named, not "traffic"

with only the first   a hobby
with only the second  a brochure
with neither          the 96.55%

This is also the fastest filter available on an existing content plan. Take ten planned pages and write both jobs for each. The ones where you cannot is not usually two or three. In my experience it is closer to half, and every one of those is a page somebody was going to write.

Almost all published content gets nothing. Ahrefs studied roughly 14 billion pages and found 96.55% receive zero organic traffic from Google, and the number is worth sitting with before reading any advice about how to write.

The band nobody quotes is the second one. 1.94% get between one and ten monthly visits, which is the band you land in when everything works: the page is indexed, it ranks for something, a person occasionally arrives, and it is worth nothing. That failure is far more common than the total failure and much harder to see, because every report on it looks fine.

I am in that band. Here is the distribution with my own reading marked on it.

Three URLs. Two free tools and a review, on a site with 11 blog posts and 32,551 words in them. Not one of those posts appears, and that is the finding rather than the embarrassment: the things ranking on this domain are the things that do something, and the things that describe things are invisible.

What do the 96.55% have in common? Ahrefs' own answer in that study is mostly about links and indexing, which is true and is steps two and five of this roadmap. The content answer is narrower and it is this: a page in that band could have been written by somebody who did not do anything. No measurement, no test, no purchase, no conversation, nothing that cost the author a morning before the writing started. It is not that those pages are badly written. Most of them are written perfectly well.

The useful thing about a number this large is that it reverses the default. The question is not "why would this page fail", it is "why would this page be one of the 3.45%". If you cannot answer that in a sentence, you have your answer.

The cheapest content decision available is the one that stops a page existing, and no content calendar has a column for it. Six questions, and any one of them can kill a row.

A page now costs almost nothing to produce and exactly as much as ever to maintain, link, update and eventually delete. That asymmetry is new, it is the defining fact of content work in 2026, and almost every process in the industry was designed before it.

1. Can I name what this contains that the top three do not? One sentence, written before the draft. If the sentence is "it is more comprehensive", the answer is no, because comprehensive is the definition of commodity.
2. Does the evidence exist yet? Not "can I find some". Does the specific artefact exist, and if not, will I actually go and produce it this week? A brief whose evidence is "research from the top 10 results" has already decided to be a retelling.
3. Is the format one I can produce? If the results page is nine calculators and I write articles, the honest outcome is not a longer article. It is either building the calculator or declining the query in writing.
4. Does the business need it? The question no results page can answer. You can satisfy an intent perfectly and gain nothing, and the pages where that happens are usually the ones with the best volume.
5. Do I already have something that does this job? Search your own site for the job rather than for the words. Half the answer to this question is an improvement to an existing page wearing a new URL.
6. Will I maintain it? Every page published is a permanent liability. If you would not update it when the interface changes, you are not publishing an asset, you are publishing a future redirect.

The sixth is the one this era needs most and the one my own audit fails hardest. Eleven posts, none of them ever updated, and I published this page anyway. The honest version of question six is not "will I maintain it in principle". It is "have I maintained anything yet", and if the answer is no then the answer to question six is no too.

A note on what killing a page actually looks like, because "just publish fewer things" is not a process. The rows you kill stay in the document, struck through, with the reason next to them. Otherwise the same row comes back in four months from somebody who does not know it was already considered, and the argument gets had twice.

Information gain is how much a document adds beyond what the reader has already seen, which makes it comparative rather than absolute. Your page is not judged on its merits. It is judged against the five things read before it.

Google filed a patent on this in October 2018, published in November 2020, titled "Contextual estimation of link information gain". Its own abstract is the clearest statement of the idea anywhere: an information gain score for a document "is indicative of additional information that is included in the document beyond information contained in documents that were previously viewed by the user."

Two things follow, and the second one is the one people miss. First, being thorough is not the same as being additive: a page containing everything the top three contain has an information gain of approximately zero by construction. Second, the comparison set is whatever the reader has already read, which means gain is contextual. Your beginner explainer has high gain for somebody who has read nothing and none at all for somebody who has read three.

The ladder below is how I use this in practice, because "add information gain" as advice produces nothing. Most content is not at level zero. It is at level one or two, which feels like work, reads well, and is still replaceable.

Level three is the line, and the useful thing about it is how close most briefs are to it. A page sitting at level one usually needs one measurement, one screenshot or one honest decision to cross over, and the reason it does not cross is never difficulty. It is that nobody asked the question before the draft started, and by the time there is a draft, the draft is what gets edited.

Now the status of the mechanism, stated plainly because this is where the industry overreaches. A patent is not a confirmed ranking system. It proves somebody at Google thought the idea was worth protecting in 2018. It does not prove anything shipped, and this page grades information gain "on the record" rather than "documented" for exactly that reason. Use the ladder because it produces better pages. Do not cite it as a ranking factor, and be suspicious of anybody who does.

Four categories of thing are unavailable to a language model and to every competitor who has not done the work: something you measured, something you ran, something you decided and regretted, and something somebody told you privately. Everything else on your page is available to everyone.

Something you measured. A number that did not exist until you produced it. This is the highest-value category and the one people assume requires scale. It does not: the corpus audit on this page covers eleven posts and took an afternoon to build a tool for, and it is the most-quotable thing here precisely because nobody else has bothered.

Something you ran. A test, a migration, a tool you paid for, a setting you changed and broke. The evidence is a screenshot with a date in the frame and the version named. Google's own guidance names a first-hand review as its example of a unique point of view, and the operative word in that phrase is not "review".

Something you decided and regretted. The most underused category in the industry, because it costs status. A decision published with its reasoning and its outcome is unforgeable: nobody can copy it without having made it, and a model cannot invent one that is true. Every unflattering number on this page is in this category.

Something somebody told you privately. A client, a vendor rep, a support ticket, a conversation at a conference. Attributable or anonymised, but real. This is the category most likely to be legitimately unpublishable, and when it is, say the finding without the source rather than dressing it as your own analysis.

What is not on that list: opinion. An opinion is free to generate and free to hold, and a strong one is not evidence of anything. The distinction I use is whether being wrong would cost me something. "I think tables are underused" costs nothing. "Three tables across eleven posts on my own site" can be checked and would be embarrassing if I had made it up.

The page is decided during research, not during writing. If you start a draft without a named piece of evidence, the draft will find one, and what it finds will be the median of the results page.

The mechanism is unglamorous. Writing is a search process: you write a sentence, you need support for it, you go and find support, and the fastest available support is whatever the top three already say. Do that for two thousand words and you have reconstructed the consensus, one honest sentence at a time. Nobody decided to write a retelling. The process wrote one.

So the order that works is: gather first, decide second, write third. Concretely, for this page, that meant reading eight competitor guides and counting their headings before writing a word, pulling two results pages and twenty keywords, and building a tool to measure my own corpus. Three artefacts existed before the first sentence, and the argument of the page came out of them rather than being illustrated by them.

Four research inputs, in the order I would spend time on them:

1. The results page, read as evidence rather than as a checklist. Not to extract subheadings. To find what every result has in common, which tells you what the commodity floor is, and what none of them has, which is your opening.
2. The forums and the reviews. Reddit and community threads contain the phrasing of the actual problem, and the complaints tell you which part of the standard answer does not work. Both results pages behind this lesson had large blocks of Reddit threads in them, and their titles were better market research than the ranking guides.
3. Your own data. Search Console, your CRM, your support inbox, your sales calls. This is where level three gain comes from and it is the input almost nobody budgets time for.
4. The primary document. If Google, or the vendor, has published on the subject, read the document rather than the coverage of it. Most of what makes this page different from the eight guides audited above is that I opened the documents they cite.

A note on using a model for this, since it is the obvious question. Research is one of the places Google's own guidance names generative AI as genuinely useful, and I use it that way: to find what has been said, build a first outline, and tell me what I have missed. It is very good at that and it is a research input, not a research output. Everything it hands you is by definition already in the commodity set.

A receipt is a claim a stranger can check. That is the whole definition, and it excludes most of what circulates in this industry as data.

The failure is not dishonesty. It is a chain: somebody measures something, somebody quotes it, somebody quotes the quote, and four hops later a number is being stated as fact by people who have never seen the method. Try it on any SEO statistic you like. Click through to the source. From there, click through again. A depressing share of them do not survive three hops, and when the chain breaks you cannot state the caveat because you never saw the method.

Two habits follow from that ladder and both are cheap. First, date every figure. A number without a date is an assertion, because the reader cannot tell whether it was true last week or in 2019. Nine of my eleven posts carry no dated figure at all, which is the count that surprised me most when I ran the tool. Second, quote the sentence rather than paraphrasing it. A paraphrase of a Google document puts you in the chain; the sentence takes you out of it, and the reader gets to argue with Google instead of with you.

The hardest habit is the third one: publish the numbers that argue against you. Every unflattering figure on this page was optional. Leaving them out would have produced a cleaner argument and a page nobody has any reason to believe, because a lesson about evidence whose every number flatters the author is itself a category of evidence, and the category is marketing.

A real named thing beats an invented person every time, and inventing a person is worse than having no example, because it tells the reader you had nothing.

"Meet Sarah, a marketing manager at a mid-sized SaaS company" is a sentence that appears in an enormous amount of content and it carries exactly zero information. Sarah does what the author needs her to do. Her problem is the problem the article solves. Her objection is the objection the article answers. No reader has ever learned anything from Sarah, and every reader who has read three of these knows what she is for.

What works instead, in order of strength. A named real thing: a tool, a site, a document, a company, with the version and the date. A number from your own data. The reader addressed directly as "you", which is honest about being a generalisation. And a real person named with permission, which is the strongest and the most expensive.

The test I use is whether the example could be wrong. "Backlinko's guide to SEO content has a Domain Rating of 90 and 13 monthly US organic visits" can be checked and I would look foolish if it were false. "A well-known SEO blog struggles to rank its own content" cannot be checked and costs nothing to say. Both sentences convey the same idea and only one of them is evidence.

The exception worth naming: a hypothetical is fine when it is labelled as one and is doing a job an example cannot. The two mocked-up first screens in chapter fourteen are invented, and they are invented because the argument is about sequence and using a real page would turn it into a critique of somebody's writing. Say "here is a made-up example" and the reader can price it correctly.

You do not need fourteen billion pages. Counting the section headings of eight competing guides took an afternoon and produced the finding this entire lesson is built on.

Original research has a reputation problem in this industry: it means a survey of a thousand marketers, or an analysis of a million SERPs, and both are out of reach for almost everybody. That reputation is doing real damage, because it stops people producing the small, specific, checkable datasets that are actually available to them and that nobody else has bothered to make.

Four kinds of very small original research, all of which I have done in the last month:

1. Count something nobody has counted. Eight guides, seventy-nine section headings, four categories. The dataset is small, the classification rule is published so you can disagree with it, and it is the only count of its kind that exists.
2. Measure your own thing with a tool you wrote. A hundred lines of regular expressions produced the corpus audit on this page. The tool is more reusable than the finding, which means the second run costs nothing.
3. Run the same input through several tools and publish the disagreement. The disagreement is the finding. Nobody publishes it because it makes every vendor look imprecise, which is exactly why it is valuable.
4. Do the thing and log it. A migration, a test, a recovery. The log is the dataset, and it only exists if you decided to keep it before you started.

The honest limit: small research is small. Eight guides is eight guides, and I would not generalise from it to the whole industry, which is why the component says the numbers are a ceiling and prints the classification rule. Stating the limit is not weakness. It is the difference between a finding and a claim, and it is what makes the finding citable.

Step four handed you a page type and a job. This step decides what goes inside it. Confusing the two is how a brief that said "comparison" produces a two-thousand-word essay with a table at the bottom.

The handover is worth being precise about, because the two steps get blurred in every workflow I have seen. Search intent decides the shape: what kind of page the results page rewards, in what format, at what depth, aimed at somebody at what stage. Content decides the substance: what specifically is in it that nothing else has. A brief that arrives with only the shape is half a brief, and the missing half is the one that decides whether the page is worth building.

from step four   comparison page · 8 of 10 results are comparisons · commercial · buyer stage
from step six    the same query run through all three tools on 23 Aug, with the disagreement

shape without substance   a well-formatted comparison that says what the vendors say
substance without shape   a genuinely new finding in a format the market does not want

Both failures are real and the second one is rarer and more painful, because you did the expensive part and then put it in the wrong container. That is chapter eighteen, and it is why format is decided by counting the results page rather than by whichever shape your team produces fastest.

One rule for the handover: the brief carries the count. Not "commercial intent" but "eight of ten results are comparisons, two are product pages, none is a guide". A label is somebody's judgement and a count is an observation, and only one of them survives being disagreed with.

The answer goes in the first screen or the page has not started. The common failure is not length, it is sequence: the answer is on the page and it arrives after the reader has already gone back.

Here is the same page twice. Same facts, same author, same total word count. The only difference is which sentence is first, and the left-hand column is not a strawman: every individual sentence in it is defensible, which is exactly why the pattern survives.

Almost every first-screen fix is a move rather than an edit. The sentence you need is usually already written, sitting at the four-hundredth word where the author finally got to the point, and the whole intervention is cutting it and pasting it above the introduction that was warming up to it.

Three things belong in a first screen and nothing else does. The answer, in a sentence. The reason to believe it, which is usually a number with a date or a source. And the shape of what follows, so somebody deciding whether to keep reading can decide. Everything else, including who you are and why the subject matters, belongs further down or nowhere.

What I will not do is attach a ranking mechanism to this. The two numbers usually quoted alongside this advice, that sixty percent of visitors never scroll and that a summary lifts conversions by a third, are both graded invented on this page because I cannot find who measured either. The reader came for an answer. That is sufficient, and it does not need a statistic to be true.

Clicking a result and immediately returning to the results page is a real behaviour and a badly documented ranking signal, and the gap between those two facts is where a lot of confident advice lives.

What is certainly true: a person who arrives, does not find the answer, and goes back has had a bad experience, and enough of those means your page is not doing its job. That statement needs no algorithm behind it and it is the reason to write a better first screen.

What is not established: that Google measures this per page and demotes you for it. Google has said repeatedly that click behaviour is not a straightforward ranking signal, that it is noisy, and that using it naively would be trivially manipulable. The industry's response has largely been to assume it happens anyway and to sell tactics for it.

My position, stated so you can discount it appropriately: I write first screens as though it matters and I do not claim it as a mechanism. That is not fence-sitting. It is the only position the evidence supports, and the practical advice is identical either way, which is usually the tell that a disputed mechanism is not worth arguing about.

Where the argument does have consequences is measurement. If you believe pogo-sticking is a ranking factor, you will chase a metric nobody can see and optimise for the appearance of engagement, which produces the intro that withholds the answer to increase time on page. That page is worse for everybody and it is a direct product of believing something undocumented with total confidence.

Google states it has no preferred word count, in a parenthesis, in a list of things that mark content as made for search engines rather than for people. The sentence is short enough to quote in full and almost nobody does.

"Are you writing to a particular word count because you've heard or read that Google has a preferred word count? (No, we don't.)"

The correlation that fuels the myth is real and it is backwards. Longer pages do tend to rank better on informational queries, because thorough answers to complicated questions take more words, and because more words gives more surface for links. Length is a symptom of having something to say, and target word counts treat the symptom.

The right unit is questions answered. A page is deep enough when somebody with the original problem can act, and the reasonable follow-up questions have been dealt with rather than deferred. That is measurable in a way word count is not: list the questions a reader will have after your first screen, and check each one is answered somewhere.

Two failure modes, and the second is much more common on good teams. Too thin is a page that answers the query and abandons the reader at the first obvious follow-up. Padded is a page that hits three thousand words by answering questions nobody asked, and it is worse than thin, because it buries the parts that were working underneath the parts that exist to hit a number.

The lesson you are reading is very long, and it is worth saying why that is not a contradiction. It is long because it is thirty-five chapters with three datasets and eight interactive drills in it, and because it has a reading-path control at the top that collapses it to seven chapters if you have twenty minutes. Length as an output of scope is fine. Length as an input to a brief is the thing Google put in a parenthesis.

Everybody scans before they read, and the scan decides whether the read happens. So the page has to work twice: once as a set of headings and bolded sentences, and once as prose.

The Nielsen Norman Group has been documenting scanning patterns since 2006, and the F-shaped pattern is the famous one: a horizontal sweep across the top, a shorter sweep further down, and a vertical scan down the left. Worth reading their own caveat too, which almost nobody quotes: it is one of several scanning patterns rather than the only one, and a page that reads well when scanned is the goal rather than a specific shape.

What survives a scan, in order of what people actually see: headings, bolded phrases, the first few words of each paragraph, list items, table rows, and anything visually distinct. What does not survive: the middle of a long paragraph, a conclusion, and anything that depends on having read the paragraph before it.

1. Headings say what the section answers. "Content decay" is a label. "Content decay is four different failures" is an answer, and somebody who reads only the headings has learned the page.
2. Bold the load-bearing sentence, not the keyword. Bolding a phrase because it is a target term trains readers to ignore your bold text, which costs you the one formatting tool that survives a scan.
3. Two or three lines per paragraph. Not because attention spans changed. Because a wall of text has no entry points and a scanner needs somewhere to land.
4. Front-load every paragraph. The first clause carries the claim and the rest supports it. A paragraph that builds to its point has hidden the point from anybody who did not read all of it.

One thing to be careful of, because this advice taken to its limit produces a bad page. Formatting is not thinking. A page chopped into eight-word paragraphs with a bold phrase in every one reads like a slide deck and cannot sustain an argument, and arguments are what a non-commodity page is made of. Scannable is a floor, not a target.

A query ending in "checker" wants an input box, and no length of article gets into that results page. Format is decided before the first sentence and it is decided by counting, not by whichever shape your team makes fastest.

This is the highest-cost mistake in the whole subject and it hides behind a process that feels rigorous. Research produces a cluster. The cluster goes into a calendar. The calendar produces a brief. The brief produces an article. At no point does anybody ask whether the market wanted an article, and step four of this roadmap has the receipt for what that costs: ten of ten results for "mortgage calculator" are calculators.

So here is the register. 22 formats, grouped by what the reader is trying to finish, with the query shape that asks for each one and what non-commodity looks like in that particular shape.

The middle column is the one that makes this a decision table rather than a taxonomy. Non-commodity means something different in each shape: for a definition page it means the circulating definitions are wrong and you can show it, for a free tool it means the tool works, and for a comparison it means criteria you chose and a verdict you are willing to be wrong about in public.

6 of the 22 have no example on this site and the gaps are left visible rather than filled with something plausible. That tells you what this content operation can currently produce, which is the same finding the exercise below will produce about yours. Almost every team makes one shape well and makes it for everything, and the queries that wanted a different shape are lost silently.

An image is carrying a claim or it is filling space, and "add images to your content" is advice that produces the second one, because it never said what the image is for.

Google's position is genuinely encouraging and it is an opportunity rather than an obligation: its generative AI features can bring in relevant images and video, which means more ways for a site to appear than a blue link. The operative word in its guidance is "relevant", and a stock photograph of somebody at a laptop is not relevant to anything.

The fourth kind, instruction, is the one with a maintenance cost attached and the one most likely to go actively wrong. A screenshot of an interface that has been redesigned twice does not merely fail to help. It sends the reader looking for a button that no longer exists, which is worse than the page having no image at all, and it is the single most common form of staleness in tutorial content.

A note on generated images, since it is now one prompt away. The rule is the same as everywhere else on this page: an illustration generated to fill a slot is a stock photograph with extra steps. Where a generated image genuinely helps is the structure category, when you need a diagram of something abstract, and there the honest version is to draw it rather than to generate a picture of a concept. Google Merchant Center already requires AI-generated product images to be labelled in their metadata, which is worth knowing about where it applies.

A table is the densest evidence format available and almost nobody builds a useful one, because the useful version requires deciding what matters and the useless version can be assembled from three marketing pages in twenty minutes.

The distinction is not presentational. A feature dump and a criteria table look identical and do opposite jobs, and the difference is entirely in what happened before the table was drawn.

The rule that does the most work is the one about deleting rows where every column agrees. A row that cannot change the reader's answer is costing them attention and buying nothing, and it survives because a longer table feels more thorough. The same instinct is what produces the twenty-row comparison where every cell is a tick.

On the claim that tables get you cited by AI systems, which you will have heard: it is graded inference on this page. It is plausible, it is repeated constantly, and I cannot find anybody who has isolated it, including the vendors selling the advice. Build the table because it lets a reader decide. That reason is sufficient, it is not disputed, and it does not stop being true if the citation claim turns out to be wrong.

My own count, from the audit: three tables across eleven posts, and nine of the eleven have none. I am arguing for the format I have most neglected, which is either the least credible position on this page or the most, depending on whether you think people should recommend things they have not done. I would rather print the number and let you decide.

A video on a page is the primary answer, or it is proof you did the thing, or it does not belong there. There is no fourth reason, and "engagement" is not one.

The primary-answer case is straightforward: some things cannot be written. Anything with timing, spatial relationships, or an interface you have to watch somebody move through. For those queries the results page usually tells you directly, because video results appear in it, and step four has the counting method.

The proof case is the interesting one and it is badly underused. A thirty-second clip of a tool actually doing the thing you claim it does is evidence in a way a paragraph is not, and it is very hard to fake casually. This is the visual-evidence category from the last chapter with a play button on it.

What does not work is a video embedded because a checklist said pages with video perform better. That correlation exists, and it exists because people who make videos are usually investing more in the page overall. Embedding a loosely related clip does not import the investment. Google's own guidance on this is measured: relevant images and video are more ways for your site to appear, and both adjectives are load-bearing.

Eight of my eleven posts embed a video and that number flatters me. Most of them are the YouTube version of the same tutorial, which is legitimate, and none of them is the thirty-second proof clip that the proof case describes. Having a video and using video are different, and the audit only measures the first one.

A section that opens on "It" or leans on "as I mentioned above" cannot be quoted anywhere, by anyone, ever. Not by a reader who skimmed to it, not by somebody sharing it, and not by a retrieval system.

The reason this happens to good writers is mechanical rather than careless. When you write a section, the previous section is on your screen, so a pronoun referring back to it is perfectly clear. It stops being clear the moment anybody arrives from a search result, a table of contents link, or a quotation, which is most of the ways sections get read.

The fix costs one sentence per section and usually costs zero extra words, because you are replacing a pronoun with the noun it was standing in for. It is the highest-return editing pass in this lesson, and the only one I would run on a page I otherwise had no time for.

I want to be precise about what this is not, because the neighbouring advice is currently being oversold. This is not chunking. Google published on 10 July 2026 that there is no requirement to break your content into tiny pieces for AI to understand it, and that there is no ideal page length. Writing sections that stand alone is a readability practice that happens to help retrieval. Cutting your page into three-hundred word blocks because somebody said models like it is a different thing, and the document those people cite says not to.

A heading promises what the next section answers. Break the promise and the reader learns to skip your headings, which costs more than any keyword the heading might have carried.

The mechanism is trust, and it decays fast. A reader who scans your headings, picks one, jumps to it, and does not find what it promised will not scan the next set. They will read linearly, get bored, and leave, which is the outcome the headings existed to prevent. Two broken headings is enough.

So the test for a heading is whether somebody who read only that heading would be able to predict the section. Three shapes that pass, and one that does not:

the question    "Which decay shape do you have?"          predicts a diagnostic
the claim       "Word count is the wrong unit"            predicts an argument
the object      "Four shapes, and what each one needs"    predicts a list of four

the label       "Content decay"                           predicts nothing at all

The label is not wrong, it is empty. It tells a scanner the subject of the section, which they could have guessed from the page title, and it tells them nothing about whether the section is the one they need. On a long page, labels are why people give up.

Two structural rules and then this chapter is finished, because heading hierarchy is on-page work and belongs to the next lesson. Use one H1, which is the title, and Google has a short video saying that more than one is not fatal. And do not skip levels for visual reasons, because the outline is the only structure a screen reader and a table of contents can see.

The check that catches everything: read your own headings alone, in order, with the body deleted. If they read as a coherent summary of the argument, the page is structured. If they read as a list of nouns, it is filed rather than structured.

"A popular SEO tool" is unrecognisable to a reader and to a machine. "Ahrefs, on the Lite plan, in August 2026" is checkable by both, and the difference costs nothing to produce.

Vagueness in content is almost never a stylistic choice. It is what you write when you do not know, when you are hedging a claim you cannot support, or when you are avoiding naming a competitor. All three are visible to a reader who does know, and the third one is visible to everybody.

What to name, specifically: tools, with the plan or version. Companies, rather than "a leading provider". Documents, with the publisher and the date it says it was updated. People, where they have said something publicly. Numbers, with the date and the source. Interfaces, with the current label of the button rather than a paraphrase of it.

There is a retrieval argument here and I will make the modest version of it. Systems that read your page work with the entities in it, and a page that says "a popular SEO tool" contains no entity to work with. That is a claim about what is legible, not about ranking, and I would make the same recommendation with no AI systems in the picture, because specificity is how a reader can tell whether you have used the thing.

What I will not do is recommend chasing mentions, and Google's July 2026 guidance is unusually direct about it: seeking inauthentic mentions across the web "isn't as helpful as it might seem". Name things because it makes your page checkable. Do not build a programme around getting named yourself.

Google published a guide to optimizing for its generative AI features on 10 July 2026, and the most useful part of it is the list of things it says you can ignore. Two of them are being sold hard by people citing that same document.

Start with what it says to do, because it is short. Create valuable, non-commodity content for your audience, which it says will "likely influence your website's presence in generative AI search in the long run more than any of the other suggestions in this guide". Keep the site technically crawlable. Add high-quality relevant images and video. Organise content so readers can follow it. That is the whole substance, and it is the first five decisions of this lesson.

Now the list of things to stop doing, quoted rather than paraphrased, because the paraphrases are the problem.

"There's no requirement to break your content into tiny pieces for AI to better understand it. Google systems are able to understand the nuance of multiple topics on a page and show the relevant piece to users."

"You don't need to create new machine readable files, AI text files, markup, or Markdown to appear in Google Search (including its generative AI capabilities), as Google Search itself doesn't use them."

And the one that most directly contradicts current practice. Google defines a fan-out query in the same document, as one of the concurrent related queries a model issues behind a single prompt, and then says this about building for them: creating separate content for every possible variation of how people might search, including fan-out queries, primarily to manipulate rankings violates the scaled content abuse spam policy, and is an ineffective long-term strategy besides.

I want to be fair to the tactic, because there is a defensible version. Looking at fan-out queries to understand what a model is actually asking is research, and research is legitimate. Generating a page for each one is the thing being named. The distinction is the same one running through this whole lesson: reading the market is fine, and manufacturing pages to match it is what the policy is about.

So the practical answer to "how do I optimize for AI search" is unsatisfying and cheap: you already know. Have something worth citing, say it clearly, name things specifically, and make sure a crawler can reach it. Every one of those is a decision earlier in this lesson, which is why this page has no chapter on GEO tactics and a chapter on refusing them.

Some queries deserve fresh results and most do not. Freshness is a property of what was asked, not of your publication date, which is why changing the date on an unchanged page is the one intervention that reliably does nothing.

The mechanism is easy to reason about from the searcher's side. "Best iphone" needs a recent answer because the answer changes annually. "How does photosynthesis work" does not, and a results page full of pages published last month would be worse rather than better. Google's own channel has a video titled "Is freshness an important signal for all sites?" and the answer is no, which is a more useful sentence than most of what is written about this.

So the first question about any page is not "when did I last update this". It is does this query deserve freshness at all, and you can answer it by looking at the dates on the current results page. If the top ten are all from the last six months, the query wants recency and you are in a maintenance race. If they span five years, recency is not what is being rewarded and a refresh cycle is spending money on nothing.

Three things that are actually freshness work, as opposed to date changes. Updating facts that have changed: prices, versions, interfaces, statistics, the law. Adding what has happened since: a new competitor, a policy change, an outcome you now know. And removing what is no longer true, which is the one that gets skipped because deleting a paragraph feels like losing something.

Two anti-patterns worth naming because both are common and both are visible to readers. Putting the year in the title and bumping it annually, which is a maintenance commitment most people do not keep and looks exactly like what it is when they do not. And displaying an updated date that changed because somebody fixed a typo, which is a small lie that costs trust the first time a returning reader notices nothing is different.

Four different failures share the word decay, and three of them are not fixed by a refresh. Telling them apart takes two minutes on one Search Console chart and it is the difference between a useful afternoon and a wasted one.

The reason this matters more than it sounds: refreshing is the default response to a falling page in almost every content team, and it is the correct response to exactly one of the four situations below. Applied to the other three it produces a page with a new date and the same problem, and it feels like progress, which is the expensive part.

Two of the four end in doing nothing, and that is the finding rather than a hedge. A page whose market has shrunk does not have a content problem, and a quarterly refresh cycle applied to it will consume hours every year for the rest of its life while the author wonders why the numbers never move.

The one most often misdiagnosed is the second. Being outclassed looks like staleness because both are gradual, and the tell is what is above you: if the new pages have something yours does not, you were beaten on substance and a date change is theatre. If you cannot name the thing they have in a single sentence, you are not ready to touch the page, because you do not yet know what you would be changing.

Update the thing that lost, not the date. Everything else about refreshing follows from that sentence, and the diagnosis from the last chapter is what tells you what lost.

Here are eight real situations and five verbs. Pick one for each before you open the reasoning, and notice how often the reflex answer is the trap.

The three ways an update goes wrong, in order of how often I see them.

1. Cosmetic. A new date, a reworded introduction, three new sentences, and nothing that addresses why the page fell. This is the most common outcome of a scheduled refresh, because a schedule produces the activity without producing the diagnosis.
2. Additive only. Two thousand words bolted onto a page that was already long enough, because adding feels safer than deleting. The page is now worse at the job it was doing, and the section that used to answer the query is four screens down.
3. Identity loss. A rewrite so thorough that the page no longer answers the query it ranked for. This one is genuinely dangerous, and the check is cheap: before you publish, open Search Console, look at what the page currently ranks for, and confirm the new version still answers the top few.

And the honest measurement problem, which is why the claim "updating recovers traffic" is graded inference on this page rather than documented. Almost every published example of a successful refresh also changed the title, the internal links and the depth at the same time. Something worked. Nobody has isolated which thing, and I have not either.

What I would do about that, practically: change one thing at a time on pages you care about, write down what you changed and when, and give it enough weeks to mean anything. Google has a video on how long SEO takes for new pages, and the same patience applies here. A refresh judged after ten days has been judged on noise.

Duplicate content is a consolidation problem with a documented mechanism. Commodity content is a value problem with no mechanism at all. They get discussed together, they look nothing alike, and only one of them is expensive.

Duplicate means several URLs answering one need, on your site or across sites. Nothing is being punished. Google picks one URL to represent the set, using signals it can see, and your preference is one input among several. The fixes are structural and they belong to step five: merge and redirect, differentiate genuinely, or canonicalise. This is a solved problem with published documentation and a short Google video.

Commodity means your page is not a duplicate of anything and adds nothing anyway. Every sentence is original prose, no other URL is competing with it, and it is still the average of the results page. There is no mechanism to explain, no canonical tag to add, and no report that flags it. It is the expensive one and it is the subject of this lesson.

duplicate   three URLs, one need        → Google picks one. Consolidate, redirect, canonicalise.
commodity   one URL, nothing new        → no penalty, no report, no traffic. Add something or delete it.
thin        any length, no added value  → Google's own framing. Word count is not the diagnosis.

Thin is the third word in this cluster and it is the most misused. Google's quality material treats thin content as content with little or no added value, at any length, and a four-thousand-word restatement of the top three is thin by that definition. The industry reads "thin" as "short" and then writes longer pages, which is how a diagnosis produces exactly the wrong treatment.

The syndication case deserves a sentence, because it is where these two genuinely overlap. Republishing your own article on Medium or LinkedIn is a duplicate question, solvable with a canonical or by accepting that one version wins. Republishing somebody else's press release, as thousands of sites do, is both: duplicate by construction and commodity by definition, which is why those pages are invisible even when the technical setup is perfect.

Google's position has not moved since February 2023: appropriate use of AI is not against the guidelines, and using automation to generate content primarily to manipulate rankings is a spam policy violation. The tool is not the line. The purpose is.

The most useful thing in that 2023 post is a historical analogy nobody quotes. Ten years earlier there were concerns about a rise in mass-produced human content, and Google's own observation is that nobody would have thought it reasonable to ban human writing in response. The answer then was to get better at rewarding quality, and it says that is the answer now.

Look at the middle band, which is three quarters of the table. Every row in it turns on the same hinge, and it is not technical: whether anybody was going to check the output. That is the next chapter, and it is why "is AI content allowed" is the wrong question. The right question is who is accountable for each fact on the page.

Where I actually use it, since you should be able to price my advice. Research and finding what has been said. Turning a transcript of me talking into structured prose, which is typing rather than thinking. First-draft sections that I then rewrite, where the rewrite always changes the facts because the draft did not have any. Alt text I check against the image. And code, which is not content.

Where I do not, and this is a preference rather than a policy: anything that constitutes the gain. If a model could write the paragraph, the paragraph is not the reason the page exists, and the paragraph that is the reason cannot be written by anything that was not there. That is not a rule I can defend from a document. It is what falls out of taking the rest of this page seriously.

One rule makes AI-assisted drafting safe: nothing ships that a human has not checked against a source they opened. Every AI content disaster is a verification failure wearing a technology costume.

The reason models are dangerous in content work is not that they lie. It is that they are fluent, and fluency reads as confidence. A wrong statistic delivered in the same register as a right one passes every quality check a human editor runs, because the quality checks were designed to catch bad writing and the writing is fine.

So the check has to be mechanical rather than editorial. Five things get verified, every time, by a person:

1. Every number. Opened, at the source, this week. Not recognised as plausible. Not remembered from somewhere. Opened.
2. Every quotation. Against the original document, word for word, including whether the person actually said it. Fabricated quotations are the highest-embarrassment failure available and they are common.
3. Every link. Clicked, and confirmed to be the thing the sentence says it is. A link to a plausible URL that does not contain the claim is worse than no link, because it manufactures the appearance of sourcing.
4. Every product claim. Against the vendor's current page, with the date noted. Pricing and features move constantly and a model's training data does not.
5. Every instruction. By doing it. If the page says click Settings then Advanced, somebody opens the product and clicks Settings then Advanced.

The uncomfortable corollary is that verification does not scale, and that is the point rather than a problem to solve. If you cannot verify a hundred pages a week, you cannot responsibly publish a hundred pages a week, and the constraint is a feature: it puts a ceiling on volume that is set by the thing that actually makes content worth reading.

Google's own guidance adds a second habit worth adopting, which is about the reader rather than about accuracy: sharing how a piece of content was created gives people useful context. That does not mean a badge on every page. It means that where automation did something substantial, saying so is more comfortable than being found out.

Scaled content abuse is defined by purpose rather than by count. Google's policy is many pages generated for the primary purpose of manipulating rankings and not helping users, and there is no number in that sentence.

Which is genuinely useful, because it means programmatic content is not the problem. A hundred thousand product pages generated from a catalog are fine, and so are location pages that contain real local information, and so is a dataset published as a page per row. Every one of those is many pages from a template and every one of them can be excellent.

The line is whether each page is worth arriving on. Three tests, and any one of them failing puts you on the wrong side:

1. Does each page contain something specific to it? Not a swapped city name. A different price, a different dataset, a different photograph, a different set of facts that somebody would want.
2. Would you be comfortable if a human had written every one? If a hundred humans writing these pages would still have produced a hundred worthless pages, the generator is not the problem and stopping using it will not help.
3. Is anybody going to check them? This is chapter thirty-one applied at scale, and it is the test that actually decides. The point at which nobody is verifying is the point at which the programme has become the thing the policy is about.

Two related policies worth knowing by name, because they catch things people do not think of as scaled content. Expired domain abuse, buying a domain for its history and repurposing it. And site reputation abuse, publishing third-party content on a host site mainly because of the host's established ranking signals, which is the policy behind the coupon sections and sponsored review directories that appeared on news sites.

The implementation of programmatic content is architecture work rather than content work, and it belongs to step five, which has the taxonomy and URL side of it. What belongs here is the decision, and the decision is the same one from chapter six asked once instead of ten thousand times: what does each of these pages contain that nobody could have guessed?

A content audit is a count of what is present, not a judgement of what is good, and that limitation is what makes it useful. A count can be repeated in six months and compared; a judgement cannot.

Almost every content audit guide starts with traffic, which is the right place to start for deciding what to prune and the wrong place for understanding what you produce. Traffic tells you which pages worked. It does not tell you that you have written eleven posts and three tables, and that fact predicts the traffic better than any per-page report will.

So here is mine, produced by a tool in the repository, on the day this page published, including the four counts that came back at or near zero.

The four counts at the bottom are the ones this lesson argues about and every one of them is bad. 9 of 11 posts contain no table, on a page with a chapter calling tables the most underproduced format on the web. 3 cite no external source at all. Nine carry no dated figure. And 11 of 11 have never been updated, on a page with three chapters on maintenance.

What a corpus audit is for, as distinct from a page audit: it tells you what your operation can produce. Mine says I produce long tutorials with screenshots, embed the matching YouTube video, cite almost nobody, and never come back. That is a coherent and quite limited machine, and no per-page report would ever have told me, because every one of those posts is individually fine.

The four counts to run on your own folder are in the exercise below, and none of them needs a tool. Posts with no table. Posts citing nobody. Posts with no dated figure. Posts never updated. Put today's date at the top and keep the file, because the second reading is the only one that means anything and it is worthless without the first.

Twelve claims I have been told about content, several of them by people who know more than me, and the ones with no source anywhere are the ones stated with the most confidence.

And now the grading, which is the part that makes the rest of this page arguable. Every load-bearing claim on it, with a grade for how much weight the evidence bears, including the four with no traceable source anywhere. Three of those four arrived in the brief for this page.

Sixteen questions. Six of them quote a sentence Google published, four are answered by measurements of this site or of the eight guides above, and none of them is a definition you could have looked up.