AI search has not changed what makes a page worth reading. It has changed what a ranking is worth. The same page that ranked third last year still ranks third; the difference is how many people now get their answer without ever seeing it.
That difference has been measured properly. In a randomised field experiment running from 7 January to 10 February 2026, researchers at the Indian School of Business and Carnegie Mellon assigned 1,065 Chrome users to see or not see AI Overviews across 68,089 searches. Where an AI Overview appeared, outbound clicks to publishers fell by 39.8% and zero-click searches rose by 34.5%.
That is a causal figure from a controlled experiment rather than a correlation drawn from traffic dashboards, which makes it the most reliable number available on this subject. It is also the whole problem in one sentence.
This guide covers what actually changed, what did not, and what is worth doing about it.
The three surfaces, and why they need separating
"AI search" is three different things with three different behaviours, and advice that does not separate them is advice you cannot act on.
| Surface | What it is | Can you measure it? |
|---|---|---|
| AI Overviews | The generated summary above Google's results | Impressions in Search Console since August 2026. No clicks |
| AI Mode | Google's conversational search interface | Impressions in Search Console. No clicks |
| External assistants | ChatGPT, Claude, Perplexity, Copilot, Gemini | Referrals when they pass one, which is often not |
Only the third sends you identifiable traffic, and it sends the least of it. The first two are where the volume is, and until very recently neither was separable at all.
That changed on 31 August 2026, when Google finished rolling out AI performance reporting in Search Console globally — impressions for AI Overviews, AI Mode and generative AI in Discover, broken out from ordinary Search. It is a genuine improvement and it stops short of what you want: impressions only, with no clicks, no click-through rate and no query data. You can now see that you appeared. You still cannot see whether anyone came.
That gap is the defining practical problem of this category, and it has its own guide: how to track AI search traffic.
What Google says you need to do
Nothing special, according to Google. Its AI features optimisation guidance, published 15 May 2026, states plainly that there is no separate markup, file or technique required for AI Overviews or AI Mode — the same crawlable, useful, well-structured pages that worked before are the input.
That is worth stating clearly because an entire cottage industry has grown around the opposite claim. The most visible example is llms.txt, a proposed file for telling language models about your site. Google has said repeatedly and specifically that it does not use it, John Mueller has compared it to the keywords meta tag, and Google's own guidance names it as unnecessary. The full position — including where the file genuinely does get used, which is not search — is in llms.txt and AI crawlers.
The honest summary: there is no AI SEO checklist that Google recognises. What exists is the ordinary discipline, applied to a surface that is much less forgiving of thin pages.
What actually changed
Four things, and only one of them is technical.
1. The value of position moved down the page
Ranking first still matters and it buys less than it did. The click that used to follow a top ranking is now sometimes absorbed by a summary that cites you without sending anyone. Being cited in an AI Overview is worth something — attribution, brand recall, occasional clicks — and it is not worth what a click was worth.
2. Informational queries were hit hardest
Queries with a short factual answer are the ones a generated summary can satisfy completely. "What is a DMARC record" is answerable in a paragraph. "Which email platform should I use given my situation" is not, because the answer depends on things the summary does not know about the reader.
The practical consequence is a portfolio decision rather than a page-level one. Content whose entire value is a definition is now competing with a free summary. Content that requires judgement, sequence, or the reader's own numbers still needs the reader to arrive.
3. Citations became a distribution channel
Assistants cite sources, and being the cited source puts your name in front of someone at the moment they are deciding. It is closer to PR than to traditional SEO, and — uncomfortably for most content teams — the things assistants cite most are frequently not your own pages. The evidence on what actually gets cited is in content formats that still get clicks from AI.
4. Being a recognisable thing started to matter more than matching a phrase
Assistants and search engines both resolve queries to entities — specific, disambiguated things — rather than to strings. A site that is clearly about something identifiable is easier to cite than one that covers everything adjacent to a keyword. That is the subject of entity SEO.
What did not change
More than the discourse suggests, and this is the section that saves money.
Crawlability and indexing. If Googlebot cannot fetch and render your page, nothing downstream happens. Broken robots rules and pages that never get indexed remain the highest-frequency, highest-cost technical failures.
Being genuinely useful. Every system in this chain is trying to identify content worth surfacing. None of them rewards a page for existing.
Internal linking and site structure. A page nothing links to is a page nothing finds — retrieval systems included.
Page speed and mobile rendering. Unchanged, and still the thing most likely to be quietly broken.
Schema, with one important correction. Structured data still helps machines parse a page. But two of the schema types most sites were told to add are now doing nothing: HowTo rich results were removed from Google in 2023, and FAQ rich results stopped appearing on 7 May 2026. The markup is still parsed; the SERP feature is gone. Anyone still selling schema work on the promise of FAQ rich results is selling a feature that no longer exists.
The measurement problem, stated honestly
You cannot fully measure this, and any tool promising otherwise is estimating.
- AI Overviews clicks arrive as ordinary organic traffic and cannot be separated in Google Analytics. Search Console now separates the impressions; the clicks remain merged.
- Google AI Mode strips the referrer deliberately, so those visits are invisible to client-side analytics entirely.
- Assistant apps on mobile — the ChatGPT app, the Copilot app — arrive as direct traffic with no referrer.
- A citation that produces no visit produces no data at all, and this is the majority case.
What you can see: AI impressions in Search Console, split by surface, page, country and device; referral traffic from assistant web interfaces that do pass a referrer; and AI crawler hits in your server logs. Between them they give you a partial, directional picture.
The most useful reading available is impressions rising while clicks stay flat — the signature of being cited without being visited, and now visible for the first time. Treat any dashboard claiming a precise "AI visibility score" as a model, not a measurement.
The setup for extracting what is genuinely available is in how to track AI search traffic in GA4.
What is actually worth doing
Six things, in order of return.
1. Answer the question in the first two sentences. Retrieval systems extract passages, not pages. A section that states its answer immediately and then explains is quotable; one that builds to a conclusion after four paragraphs of preamble is not. This is the single highest-return structural change available, and it improves the page for human readers at the same time.
2. Make every section self-contained. Assume the reader — or the model — arrives at that heading and reads nothing before it. Backward references like "as we saw above" break the passage the moment it is lifted out of context.
3. Publish things that cannot be summarised away. Original data, your own test results, specific procedures with real settings, judgement about trade-offs. A summary can replace your definition. It cannot replace your evidence.
4. Be specific and checkable. Dates, figures, version numbers, named error codes. Specifics are what make a source worth citing rather than paraphrasing, and they are what a fact-checking pass can confirm.
5. Fix your entity basics. Consistent organisation naming, an About page that says plainly who you are, sameAs links to your profiles elsewhere, Organization and Person schema. Cheap, one-off, and it makes you a resolvable thing rather than a string.
6. Decide about crawlers and the AI opt-out deliberately. Blocking the wrong bot removes you from AI answers today while doing nothing about model training. Google also now offers a first-party control to opt out of its generative AI features entirely — which forfeits the impressions along with the traffic. Both decisions are in llms.txt and AI crawlers.
What is not worth doing
llms.txt, for search visibility. Google does not use it. It has a real use, and that use is not this.- Adding FAQ or HowTo schema to chase rich results. Both features are retired. Keep the markup where it honestly describes the page; do not commission work to add it.
- "AI-optimised" content spun to include more question phrasings. The systems are summarising meaning, not matching strings.
- Paying for an AI visibility score before you have checked what your server logs already tell you for free.
- Rewriting a working site around a technique announced three weeks ago. This field's half-life is short, which is exactly why the durable advice is unglamorous.
Frequently asked questions
Is SEO dead because of AI?
No, though the value of a ranking has fallen measurably. A controlled experiment across 68,089 searches found that when an AI Overview appeared, outbound clicks to publishers fell 39.8% and zero-click searches rose 34.5%. Ranking still determines whether you are eligible to be cited or clicked; it now converts to traffic at a lower rate.
What is AI search optimisation?
Making pages that generative search systems can retrieve, understand and cite. In practice it is ordinary technical and editorial quality — crawlability, clear structure, answer-first passages, specific checkable claims — applied to a surface that is less forgiving of thin content. Google states no special markup or file is required.
Do I need an llms.txt file?
Not for search visibility. Google has said specifically that it does not use llms.txt, and its May 2026 AI optimisation guidance names it as unnecessary. The file does have genuine uses in agent and developer workflows, which is a different problem from being found in search.
How much traffic do AI Overviews take?
The best available figure is a 39.8% reduction in outbound clicks on searches where an Overview appears, from a randomised experiment rather than an observational study. It varies enormously by query type — informational queries with short factual answers lose most, and queries requiring judgement or the reader's own context lose least.
Can I track traffic from ChatGPT and Perplexity?
Partly. Their web interfaces often pass a referrer you can capture in analytics, but their mobile apps strip it and arrive as direct traffic. Google AI Mode removes the referrer deliberately, and AI Overview clicks are indistinguishable from ordinary organic. Any complete-looking AI traffic number is an estimate.
Does schema markup help with AI search?
It helps machines parse a page, which is worth having. But two schema types widely recommended for this purpose no longer produce anything in Google: HowTo rich results were withdrawn in 2023 and FAQ rich results stopped appearing on 7 May 2026. Keep markup that honestly describes your content; do not add it expecting a retired feature.
Should I block AI crawlers?
Only after separating them, because they do different jobs. Training crawlers feed future models; retrieval crawlers fetch pages so an assistant can cite you in an answer right now. Blocking the retrieval ones removes you from AI answers today, which is usually the opposite of what people intend.
What content still earns clicks in AI search?
Content a summary cannot replace: original data, your own test results, specific procedures with real settings, and judgement about trade-offs that depend on the reader's situation. Definitional content is the most exposed, because a definition is exactly what a generated answer does well.
What to do next
Open Search Console and look at the AI performance report — it rolled out globally on 31 August 2026, so there may be less history than you expect. Compare your AI impressions against your overall Search clicks for the same pages. That comparison tells you more about your exposure than any third-party AI visibility tool, and it is free.
Then take your five highest-traffic informational pages and check whether each one answers its question in the first two sentences. Most do not, and that is the cheapest fix available.
Related guides
- Entity SEO: why named things beat keywords now — becoming a resolvable thing rather than a string
- llms.txt and AI crawlers: block them or not? — the file, the bots, and the decision
- How to track AI search traffic in GA4 — what is measurable and what is not
- Content formats that still get clicks from AI — what actually gets cited
- Does Google penalise AI content? — the production-side question
- Fact-checking AI content before you publish — why specifics are the asset
Chapters in this Guide
Free: The 60-Minute Email Authentication Fix
A no-fluff checklist to set up SPF, DKIM & DMARC correctly and pass Gmail & Yahoo's sender requirements.
Muhammad Basim has worked in digital marketing since 2013, focused on email deliverability and AI-assisted content production. He is the author of The Email Deliverability Playbook and The Email Copywriting Playbook.
Related Articles
Email Deliverability: Why Authenticated Emails Still Land in Spam
Email deliverability is whether your message reaches the inbox rather than the spam folder, and it is decided by three factors: technical infrastructure, list quality, and sending behaviour. Authentication is one item inside the first factor. Getting it right is necessary, and it is nowhere near sufficient. That gap explains the most common complaint in […]
Linkable Assets: What Actually Earns Links
A link is a citation, and a citation requires that someone writing about your subject needed you to make their point. That is the whole mechanism, and it is the reason most "linkable content" earns nothing. The test, before you build anything: Could a writer covering this topic finish their sentence without referencing you? If […]
Link Outreach Emails That Get Replies
Outreach fails for exactly two reasons, and only one of them gets written about. The first is that the email was not worth replying to. That is the reason every guide addresses, and the advice — personalise, be brief, offer value — is correct and insufficient. The second is that the email never arrived. Nobody […]