AI search
Google is right about Google
Google’s documentation says optimizing for AI search is “still SEO”1, and inside Google’s own products the data agrees: AI Overviews still take 38% of their citations from pages ranking in the top ten3. But that guidance describes AI Overviews and AI Mode in Google Search, and says nothing about ChatGPT, Copilot, Perplexity or Claude2. Outside Google, ChatGPT’s citations overlap Google’s top ten about 8% of the time4.
That gap is worth understanding, and it is not worth a second retainer. This work is part of the SEO engagement, not a service sold beside it. This page exists to set out exactly where the two differ.
Every number here is linked to the study it came from, with its date and sample size. Some of those sources cut against the argument being made, and they are marked.
The opposing case, first
What Google actually says
“From Google Search’s perspective, optimizing for generative AI search is optimizing for the search experience, and thus still SEO.”
They go further. You do not need llms.txt or AI text files, you do not need to break content into chunks, you do not need to write differently for AI, and there is no special schema markup to add1. Every one of those is something the AI SEO industry currently sells. Google says all four are unnecessary, and on its own features Google is the authority.
This page agrees with all of it. The disagreement is one word long, and it is a word Google put in the sentence itself: perspective.
Where it stops holding
How often an AI’s citations also rank in Google’s top ten
If AI visibility were simply a by-product of ranking, these numbers would be high and they would be similar. They are neither.
| Platform | In Google’s top 10 | Notes |
|---|---|---|
| Google AI Overviews | 38%3URL-level | Down from 76% in July 2025. Google's own feature, and still the closest to Google's rankings by a wide margin. |
| Perplexity | 28.6%4URL-level | The nearest thing to a search engine among the assistants, and it shows. |
| Claude | 12.6%13Domain-level | Measured on the more generous of the two metrics and still low. Claude does not appear to query Google at all: its web search runs on Brave, per Anthropic's own subprocessor list. |
| Gemini | 8.6%4URL-level | Google's own model, used as an assistant rather than inside Search, behaves like the assistants. |
| Copilot | 8.2%4URL-level | Built on Bing, and overlapping Bing's own top ten only 16.6% of the time. |
| ChatGPT | 8.0%4URL-level | More than 80% of its citations come from pages that do not rank for the query at all. |
These are three different studies and should not be read as one scale. The AI Overview figure comes from 863,000 keyword SERPs and four million citations analysed in March 20263. The ChatGPT, Gemini, Copilot and Perplexity figures come from a separate run of 15,000 long-tail queries in August 20254. A later test on 3,311 short-tail queries put ChatGPT at 10% overlap at URL level and 31.8% at domain level5, so the exact number moves with query type and with the month.
Claude is measured differently, and the difference favours it. No URL-level study covers Claude — the cross-platform run above left it out — so its figure comes from a peer-reviewed analysis of 1,000 ranking queries that measured overlap at domain level13. Domain level is the looser test: any page on a domain Google ranks counts as a hit. On that same consistent measure the paper put Perplexity at 15.2%, Claude at 12.6%, Gemini at 11.1% and GPT-4o at 4.0%, with GPT-4o’s median overlap at zero — for more than half of its queries, not one cited domain appeared in Google’s top ten13.
Claude is the clearest case of the whole argument. It does not appear to consult Google at all. Its web search runs on Brave, which TechCrunch identified from Anthropic’s own subprocessor list, a BraveSearchParams parameter in the API, and identical citations returned by both14. Anthropic declined to comment, so treat it as well-evidenced reporting rather than confirmation. If it holds, then for one of the fastest-growing assistants the index your SEO work targets is not the index being read.
What does not move is the shape. Google’s own feature tracks Google’s rankings; the independent assistants substantially do not. And they do not agree with each other either — 86% of the fifty most-mentioned sources are not shared across ChatGPT, Perplexity and AI Overviews6. The analysis of ChatGPT concludes it “doesn’t just lift URLs directly from Google, Bing, or other indexes” but applies its own re-ranking on top5.
Which of these matters depends on what you measure
By share of generative AI website visits, Claude reached about 9% worldwide in May 2026, up from roughly 2% a year earlier and the fastest-growing platform in the category15. On that measure it is comfortably ahead of both Perplexity and Copilot.
By share of referral traffic actually arriving at websites, the order reverses. StatCounter had Claude at 2.57% in August 2026, behind Perplexity at 4.31% and Copilot at 2.79%16.
Both are true, and the gap between them is this page’s argument compressed into a single platform. Claude is used heavily and links out rarely — the Toronto researchers noted it returned no links at all for most informational and transactional queries unless explicitly prompted to search13. It is a large, fast-growing audience that will almost never appear in your analytics. Judge these platforms by the traffic they send and you will conclude Claude does not matter; you will be measuring the wrong thing.
What actually moves it
Five levers, ordered by how well evidenced they are
Not by how good they sound. The last one is a null result and is on the list for exactly that reason.
Being quotable, in the specific way the research means
The peer-reviewed work on this tested nine interventions against a 10,000-query benchmark. Adding relevant quotations, statistics and citations to a page moved its visibility in generative answers the most; keyword stuffing, the classic search tactic, produced almost nothing. The authors' own conclusion is that techniques effective in search engines may not transfer to this.
Evidence: Quotations +41%, statistics +33%, citing sources +28% on the paper's visibility metric. Keyword stuffing: little to no improvement.7
Being talked about somewhere that is not your website
The largest correlation study on this looked at 75,000 brands and found that what tracks with being mentioned by an AI is being mentioned by people. Branded web mentions and YouTube mentions sat at the top of the table. Backlinks and URL rating, the two numbers most SEO retainers are sold on, sat at the bottom of it.
Evidence: Branded web mentions 0.656–0.709; YouTube mentions ~0.737; backlinks 0.141–0.222; URL rating 0.125–0.197. Correlation, not causation, as the authors say.8
Being unambiguous about who you are
A model has to resolve you to an entity before it can recommend you, and it assembles that from everything it has seen rather than from your homepage. One name, one description, the same facts wherever they appear. This is the least glamorous item here, and the one whose evidence is the most indirect.
Evidence: Inferred from the mention correlations rather than tested directly. Flagged as the weakest-evidenced item on this list.8,6
Ranking, which still matters and is not sufficient
Sites with more organic traffic are mentioned more, Google's AI features still pull disproportionately from the top ten, and anyone telling you ranking is irrelevant is overselling. It is the floor rather than the mechanism: necessary often, sufficient rarely, and close to absent as a predictor on ChatGPT.
Evidence: Organic traffic against AI mentions: Perplexity 0.66, AI Overviews 0.47, ChatGPT 0.33 — and ChatGPT's result is not significant at the 5% level (p=0.0515).10,3
Schema, which did not work when it was tested
Worth having for search generally, and the one controlled study of its effect on AI citations found essentially nothing. It is on this list because it is the single thing this industry sells hardest, and because a page claiming to be evidence-led has to report the results that were inconvenient as well as the ones that were not.
Evidence: 1,885 pages adding JSON-LD against 4,000 matched controls: AI Mode +2.4%, ChatGPT +2.2% — both statistically indistinguishable from zero — and AI Overviews −4.6%.9
The distinction
AI SEO and SEO, line by line
Six rows differ and four are identical. That ratio is the argument: different enough to need naming, not different enough to be a separate discipline or a separate invoice.
| Aspect | SEO | AI search |
|---|---|---|
| What you are trying to winDiffers | A position in a ranked list of links | A sentence inside a generated answer |
| Who decidesDiffers | One ranking system | Several, which demonstrably disagree with each other |
| Which index is readDiffers | Google's | Google's, Bing's, Brave's and the assistants' own re-ranking |
| How you measure itDiffers | Rank tracking and Search Console | Asking the models directly, per platform, and recording the answers |
| What winning gets youDiffers | A click | Usually a mention with no click at all |
| Levers unique to itDiffers | Keyword targeting, internal linking, click-through rate | Quotability, entity clarity, being mentioned off your own site |
| Technical foundationsSame | Crawlable, fast, indexable, structured | Identical. There is no second technical standard |
| Content qualitySame | Complete, accurate, genuinely useful | Identical, and for the same reasons |
| RankingSame | The whole objective | The floor. Helps everywhere, decides nothing outside Google |
| How it is bought hereSame | A monthly engagement | Part of that engagement, not a second one |
So this is not sold as its own retainer
Four of the ten rows above are identical, and they include the ones that take the most work: the technical foundation, the content, the ranking underneath it. Charging a second monthly fee for a second pass over the same ground would be hard to defend, and the common industry practice of billing AI search separately is the thing this page is least comfortable with.
So the work in the differing rows — the per-platform baseline, the entity cleanup, the quotability pass, the off-site mentions — is part of the organic SEO engagement, inside the same $1,000 to $5,000 a month. The only thing quoted separately is the one-off baseline audit, because it is genuinely useful on its own and some people only want that.
Against our own argument
Three things that cut the other way
Ranking and AI visibility really are connected. Sites with more organic search traffic are mentioned more by AI systems — correlating 0.66 on Perplexity, 0.47 on AI Overviews and 0.33 on ChatGPT10. Anyone telling you rankings no longer matter is selling something.
The researchers whose data this page leans on disagree with part of it. Their own conclusion, in the study showing the platforms cite different sources, is that “good content is good content, and top websites are top websites. These things don’t change.”6 That is a fair reading of their data, and it is quoted here rather than left out.
Almost everything above is correlation. The 75,000-brand study states plainly that correlation is not causation and that improving these metrics will not automatically improve AI visibility8. The one genuinely controlled experiment cited on this page is the schema study, and its result was that the intervention did nothing9.
The argument survives all three, but it survives as something narrower than the industry claims: the same foundations, a different target, measurement a rank tracker cannot do, and a handful of levers classic SEO does not pull. Not a separate discipline. Not a secret checklist. A different job done with much of the same equipment.
The commercial case
Small volume, unusual value, and a cost you pay either way
8%
of searches produced a click on a traditional result when an AI summary was present, against 15% without one. Pew Research, 68,879 searches.11
1%
of visits involved clicking a source link inside the AI summary itself. Being cited is mostly visibility rather than traffic, which is the honest frame for all of this.11
4.4×
the value of an average organic visit, for the AI search visitors who do arrive — on a study scoped to digital marketing topics, with extrapolated projections.12
Read together those say something more useful than any of them alone. The traffic is small, and the first two numbers are a cost you incur whether or not you do anything about it: when an AI answer appears, most of the click is gone already. The work is a hedge on where attention is moving, with a decent expected value and no guarantee. Anyone presenting it as a growth channel for this quarter is ahead of the evidence.
Inside the SEO engagement
What the work consists of
These are the tasks the differing rows above actually produce. They run inside the organic SEO engagement, alongside the ranking work rather than instead of it, and at no separate monthly fee.
If your foundations are not in place yet, none of this is where to start. Ranking, consistent listings and being mentioned anywhere at all are the same work seen from a different angle, and they pay off whatever the models do next.
- A baseline of what each model says about your category now, recorded per platform because they disagree
- Entity cleanup, so your site, profile and listings stop contradicting each other
- The pages a model would need to quote, written so that they can be
- Work on being mentioned off your own site, which is what the correlations actually point at
- Structured data, done properly and sold as hygiene rather than as the mechanism
- Re-checks on a schedule, with the months where nothing changed reported as months where nothing changed
Reasonable doubts
The questions a sceptic asks
Mostly we are not. Google's guidance says optimizing for generative AI search is still SEO, and inside Google's own products the data broadly agrees: AI Overviews still take about 38% of their citations from pages ranking in the top ten. The disagreement is about scope. Both of Google's documents describe AI Overviews and AI Mode in Google Search, and neither makes any claim about ChatGPT, Copilot, Perplexity or Claude. Outside Google the correspondence largely goes away — ChatGPT's citations overlap Google's top ten around 8% of the time. Google is right about Google and silent about the rest, and the rest is where the work differs.
Fair question, and the honest answer is that a lot of what is sold under this name is an upsell. There is no secret checklist. Google has explicitly said llms.txt, content chunking, AI-specific rewriting and special schema markup are not needed for its features, and the one controlled test of schema found no effect on citations. What survives is narrower: the same foundations, a different target, measurement a rank tracker cannot do, and a small number of levers — quotability, off-site mentions, entity clarity — that the peer-reviewed work found move generative visibility while classic tactics like keyword stuffing did not. That is worth doing deliberately. It is not worth a second retainer, which is why it is not sold as one here.
Yes, and this page would be dishonest if it said otherwise. Sites with more organic search traffic get mentioned more by AI systems, and the researchers behind much of the data cited here conclude that good content is good content and top websites are top websites. The qualification is that the relationship is moderate rather than mechanical — correlating 0.66 on Perplexity, 0.47 on AI Overviews, and 0.33 on ChatGPT, where it is not statistically significant. Ranking is the floor. It is not the whole mechanism, and on the largest assistant it is close to not being a predictor at all.
Because there is no rank to track. AI Overviews are probabilistic and their content changes with every query, by the description of the researchers who assembled the largest dataset on them. The platforms also disagree with each other: 86% of the fifty most-mentioned sources are not shared across ChatGPT, Perplexity and AI Overviews, so a single visibility score averaged across platforms hides more than it reports. What can be done is direct observation — ask the models the questions your customers ask, per platform, on a schedule, and record what comes back.
Not yet in volume, and possibly in value. AI referrals are a low single-digit share of traffic for most small businesses today. But Pew Research found that when an AI summary appears, clicks on traditional results roughly halve — 8% against 15% — and only 1% of visits involved clicking a link inside the summary itself, so the visibility you lose there is real whether or not you pursue the citation. Semrush put the average AI search visitor at 4.4 times the value of an organic one on conversion rate, though that study was scoped to digital marketing topics and its projections are extrapolations. Treat it as a hedge with a decent expected value, not a channel that will replace your pipeline this year.
No, and nobody can. The researchers holding the largest public dataset on this describe AI Overview citations as probabilistic and changing with every query, and their analysis of ChatGPT concludes there are additional factors influencing which URLs get surfaced that sit outside a publisher's control. When the people with the data will not claim a formula, be careful with anybody who does.
No. It is part of the SEO engagement and there is no separate fee for it. The comparison on this page sets out where the two differ — the target, the measurement, the platforms, and three levers classic SEO does not pull — and where they are identical, which is the technical foundation, the content and the ranking underneath. The differences are real enough to be worth understanding before you buy anything, which is why this page exists. They are not large enough to justify billing you twice.
Nothing on top of SEO. This work is part of the organic SEO engagement, which runs $1,000 to $5,000 a month depending on the market and the state of the site. Much of the industry sells AI search as a second monthly retainer alongside SEO; given how much of the underlying work is identical work, that is difficult to justify. The one thing quoted separately is the baseline audit — what each model says about you now, where your entity contradicts itself, and which pages are closest to being quotable — priced as a one-off from the size of the site, because it is useful on its own and some people only want that.
Sources
Every claim above, and where to check it
Dates and sample sizes are shown because they matter: a study of 900 people and a study of four million URLs are not the same kind of evidence. Sources marked in ochre either weaken the case made here or state the opposing position.
- 1Google's guide to optimizing for generative AI features on Google Search
Google Search Central · Updated 10 July 2026 · Official documentation
The position being argued with
- 2AI features and your website
Google Search Central · Documentation · Official documentation
The position being argued with
- 3Update: 38% of AI Overview citations pull from the top 10
Ahrefs · 2 March 2026 · 863,000 keyword SERPs, 4 million AI Overview URLs
- 4Only 12% of AI cited URLs rank in Google's top 10 for the original prompt
Ahrefs · 11 August 2025 · 15,000 long-tail queries across four assistants
- 5ChatGPT may scrape Google, but the results don't match
Ahrefs · 3 September 2025 · 3,311 short-tail queries
- 686% of top mentioned sources are not shared across ChatGPT, Perplexity and AI Overviews
Ahrefs · 12 June 2025 · 76.7M AI Overviews, 957k ChatGPT and 953.5k Perplexity prompts
- 7GEO: Generative Engine Optimization
Aggarwal et al., KDD 2024 (Princeton, Georgia Tech, AI2, IIT Delhi) · Peer reviewed, 2024 · GEO-bench, 10,000 queries across nine domains
- 8Top brand visibility factors in ChatGPT, AI Mode and AI Overviews
Ahrefs · 12 December 2025 · 75,000 brands, Spearman correlations
- 9Does schema markup affect AI citations?
Ahrefs · 11 May 2026 · 1,885 pages adding JSON-LD against 4,000 matched controls
- 10Websites with more organic search traffic get mentioned more in AI search
Ahrefs · 23 June 2025 · Top 50 mentioned sites across three AI systems
Cited against this argument
- 11Google users are less likely to click on links when an AI summary appears in the results
Pew Research Center · 22 July 2025 · 900 US adults, 68,879 searches, 12,593 with AI summaries
- 12We studied the impact of AI search on SEO traffic
Semrush · 21 July 2025 · 500+ digital marketing topics; projections are extrapolated
- 13Navigating the Shift: A Comparative Analysis of Web Search and Generative AI Response Generation
Chen, Wang, Chen & Koudas, EDBT/ICDT 2026 Workshops · Peer reviewed, revised May 2026 · 1,000 ranking queries across ten consumer topics, domain-level Jaccard overlap
- 14Anthropic appears to be using Brave to power web search for its Claude chatbot
TechCrunch · 21 March 2025 · Reporting on Anthropic's subprocessor list; Anthropic declined to comment
- 15AI search stats in 2026
Similarweb · 29 July 2026 · Generative AI website visits, worldwide, May 2026
- 16AI chatbot market share
StatCounter Global Stats · August 2026 · Referral traffic reaching sites in StatCounter's network
Cited against this argument
A weakness worth naming: most published research in this field comes from SEO tool vendors who also sell AI visibility products, and 7 of the 16 sources here are one such vendor (Ahrefs). That is a real limitation of the evidence base rather than a choice made on this page. The two strongest sources are not vendors — Pew Research is independent, and the GEO paper is peer reviewed at KDD — and where a vendor study is doing the work, its sample and method are stated above so you can weigh it yourself.
Last reviewed September 2026. This field moves quarterly, and a page like this one that has not been touched in a year should be read with that in mind — including this one.
Start by finding out where you stand
Name your category and the questions your customers ask, and you will get back what each model actually says today — per platform, because they disagree, and including if the answer is somebody else. That baseline is a one-off; everything after it runs inside the SEO engagement.