LLM SEO 2026: we asked four AI engines the same 40 buying questions.

How this was measured

There is no official statistic for what AI assistants cite. There are vendor blogs, and there are tools that sell you a score. So I ran the thing itself.

40 questions, in five groups of eight, all of them questions a real buyer types before spending money: who to hire, what it costs, which tool to use, how to do something, and one thing versus another. They sit in the service-business space — agencies, video, branding, websites, social, SEO — because that is the market I can check the answers against.

Four engines, web search enabled, US market, one live call per engine per question on 15 August 2026: ChatGPT (gpt-5.4-mini), Perplexity (sonar-pro), Claude (claude-sonnet-4-5), Gemini (gemini-3.5-flash). That is 160 answers, with every annotated source recorded. Then Google's own organic top 20 for the same 40 questions as a control — 37 of 40 returned a result set.

Then a second pass on the infrastructure: all 902 domains that any engine cited or Google ranked were checked for robots.txt and llms.txt — and, as a control group, 980 ordinary German local-business websites (hairdressers, bakeries, workshops) collected for an earlier study, checked the same way on the same day.

What these numbers do not say

Four limits, which I would rather state myself. First: this is one snapshot on one day, in one market. AI answers are not stable — ask again next week and the sources move. Second: one model per engine. A bigger or smaller model in the same family may cite differently, and the assistant your buyers actually use may not be the one tested here.

Third: a citation is not a recommendation. Being named as a source is not the same as being named as the answer, and this study counts sources, not praise. Fourth: an engine can lean on something it never annotates. What is measured here is what each engine was willing to show as its source — that is the only part anybody outside the lab can verify.

What the numbers do say: how wide the disagreement between engines is, what kind of page gets used, and how much of it Google's ranking already predicts. That is enough to decide what to do on Monday.

Four engines, four different worlds

Start with volume, because the engines are not remotely comparable. Asked the same question, one gives you a shortlist of two sources and another gives you eighteen.

That single chart reframes the whole exercise. If your buyers use ChatGPT, you are competing for roughly two slots. If they use Perplexity, there are eighteen, and being one of them means much less. Any vendor selling you one "AI visibility score" is averaging over four markets with different rules.

Now the disagreement. For each question I compared which domains each pair of engines cited:

Read the bottom row again: the two assistants most people actually use agree on 4.3 percent of their sources. Across 31 questions where all four engines cited something, the mean number of domains shared by all four was 0.26 — and only four questions produced even one such domain.

The long tail says the same thing from the other side. Of 699 domains, 532 — 76.1 percent — were cited exactly once in the entire study. The ten most-cited domains together account for just 14 percent of all citations. There is no small club that owns AI answers in this category.

This is the good news, and it is worth stating plainly, because most coverage of AI search is written as if the door were closing. A market where three quarters of cited sources appear once is a market with room in it. It is also a market where nobody can promise you a position.

Nobody searches your question

Two of the four engines expose the search queries they ran before answering. Across 82 answers where those queries were visible, here is the count that matters:

Instead they take the question apart and rebuild it — 2.67 queries per answer on average, up to eight. Ask "Which agency should I hire to produce a corporate explainer video?" and ChatGPT goes looking for best corporate explainer video agency 2026 explainer video production agency corporate clients, then corporate explainer video agency portfolio motion graphics corporate video agency.

That alone would only be a keyword lesson. The second half is the one nobody seems to have written down.

The engine finds candidates from third-party sources, then walks to each candidate's own site to check them. In other runs it does this with an explicit site: operator — site:sculpt.co B2B social media marketing LinkedIn official. The word official shows up again and again: it is looking for the company's own statement of what it does.

This is the step where most websites lose. You can be mentioned in a listicle, land on the candidate list, get visited — and then fail the check, because the page the engine lands on says "we deliver bespoke solutions that drive impact" instead of what you actually do, for whom, at what price, with what result.

What that means for the page you own

The verification query is a fact-matching operation. It succeeds when your page states, in plain text, the things the query asks about: the service in the words buyers use, the segment you serve, the platform or method named, evidence, and — the most avoided of all — the price. Three of the five question groups in this study (price, tool choice, comparison) are answered almost entirely from pages that state specifics.

There is a second-order effect worth planning for: because the engine writes its own queries, it also writes queries you never targeted. You cannot rank for a rewrite you have not seen. What you can do is make sure the page carries the underlying facts in several plain formulations rather than one polished slogan.

What kind of page gets cited

The received wisdom is that AI answers are stitched together from Reddit threads and review directories, and that a normal business site has no chance. Sorted by category, the 1,121 citations look like this:

Reddit is the single most-cited domain — and it is still under four percent of everything. The picture is not a handful of gatekeepers; it is a very long tail of ordinary websites, most of them named once.

The one place where the directories do concentrate is exactly where you would expect: "who should I hire" questions. There, directories take 14.0 percent of citations, four times their share elsewhere. On price questions it drops to 3.8 percent, on comparisons to 1.6 percent.

So the practical split is: for vendor-choice questions, a profile on the two or three directories in your category is worth having, because that is where candidate lists get built. For everything else — cost, method, comparison, how-to — the citation goes to whoever actually wrote the specifics down, and that can be you.

How much of this Google already decides

For the same 40 questions I pulled Google's organic results, then compared them against what the engines cited.

Two numbers, and they point in opposite directions. Averaged per question, 27.7 percent of the domains cited by AI also appear in Google's top 10. So seven of ten AI citations come from outside the first page. But looked at the other way, 67.2 percent of Google's top 10 gets cited by at least one engine — and there was not a single question with zero overlap.

Both are true and both matter. Ranking on page one makes you a likely candidate for citation. It is nowhere near sufficient, and it is not the only door: most of what gets cited was not on page one at all.

And the number that should end the "AI versus Google" framing entirely: Google returned an AI Overview on 37 of the 37 questions that produced results. Every single one. The biggest generative answer engine in this study was not ChatGPT — it was the search engine that already sends most of your traffic.

Which is why I am careful with the traffic argument. AI assistants still send well under one percent of referral traffic to publishers, while Google sends around 88 percent of search referrals. If someone tells you to rebuild your site because AI traffic is about to replace search traffic, the numbers do not support them. The reason to care is not the traffic AI sends. It is whether you exist in the answer at the moment somebody decides who to call.

Can the engine even read your site?

Every domain in this study got a second check: what its robots.txt says to AI crawlers, and whether it publishes an llms.txt. Then the same check on a control group — 980 ordinary German local-business websites, the kind of company that is not thinking about AI search at all.

The headline difference is not that one group blocks and the other does not. It is that one group has decided anything at all. Cited sites name GPTBot in their robots.txt eleven times as often, and publish an llms.txt more than four times as often.

Now the part that costs ordinary businesses money. OpenAI runs more than one crawler, and they do different jobs. From OpenAI's own documentation: GPTBot crawls content that may be used to train models. OAI-SearchBot is the one that, in OpenAI's words, is "used to surface websites in search results in ChatGPT's search features". ChatGPT-User is a fetch triggered by a person.

Read those two pairs together. Sites that get cited block the training crawler more often and the search crawler less often. Ordinary business sites do the exact opposite: they are twice as likely to block the one crawler that decides whether they appear in ChatGPT's search results at all.

Nobody chose that. It is what a blanket Disallow from a security plugin, a hosting default or a copied template does when it meets a crawler nobody in the company has heard of. Two other numbers from the same scan point the same way: 18.9 percent of the local-business sites have no robots.txt at all, and 14.4 percent did not answer within twelve seconds.

About llms.txt, honestly

llms.txt is a markdown file at your site root that hands an AI agent a clean summary and a map of your important pages. Jeremy Howard proposed it in September 2024; version 2 was published on 10 August 2026, OpenAI, Anthropic and Google all publish one for their own documentation, and Chrome's Lighthouse now audits for it.

And no engine has confirmed that it affects citation. So I will not tell you it is a ranking factor — that claim is currently unsupported. What the data shows is narrower and still interesting: the sites that get cited are four and a half times more likely to have one. That is a correlation with the kind of team that also does the other things right, not proof that the file did the work. It costs an hour. It cannot hurt. That is the whole case for it.

The mistake I found on my own site

This site is built for exactly what this article describes. AI crawlers are explicitly invited in robots.txt. There is an llms.txt. There is FAQPage markup on the articles. I have been writing about generative search since last year.

On 15 August 2026 I ran the check on heidarrudyi.com/cases — the page that holds every piece of work I have to show. It contained no static link to a single case study. The whole grid was drawn from a JavaScript array at load time.

Google renders JavaScript, so Google saw the portfolio. The AI crawlers I had invited in largely do not render JavaScript. They saw an empty page. A site optimized for AI visibility was hiding its fifteen pieces of proof from the machines it had personally invited.

The fix took twenty minutes: a visible text index of all fifteen links underneath the grid — useful to people as quick navigation, and real links for anything that cannot run JavaScript. Not hidden text, because hidden text is a different problem.

I mention this because it is the most common version of the problem, and it has nothing to do with llms.txt, schema markup or any AI-specific tactic. The content simply was not there. Before optimizing anything for AI, check that a machine can see what you already have.

What I would actually do on Monday

Step four: pick your two directories, then stop. Directories take 14 percent of citations on „who should I hire" questions and almost nothing everywhere else. A profile on the two that matter in your category is worth an afternoon. A profile on twenty is a subscription you will not notice paying.

Step five: publish the specifics nobody else will. Three of the five question groups here — price, tool choice, comparison — are answered from pages that state numbers. Most companies in most industries refuse to publish prices, methods and outcomes, which means whoever does becomes the only citable source in the category. That is also why my own prices sit openly on one page rather than behind a call.

What I would not do: buy an „AI visibility score". This study measured four engines that overlap by four to fourteen percent, on one day, with results that move. A single number averaged across them tells you nothing you can act on.