Reference guide

What is AEO? Answer engine optimization, explained with data

A TofuBofu reference guide, maintained from our own measurement.

Most definitions of AEO are written by people who have not measured it. This one is not. Every claim below is either a plain definition or a number, and every number is either one we produced by asking real engines real buying questions, or one published by a named source you can go and check.

Where the published studies disagree with each other, which happens more than the category admits, this article says so and explains why rather than picking the convenient one.

The evidence base for this article

So you can weigh each claim, here is what everything below is built on. Four of these are ours and open to inspection, four are third-party and linked.

What an answer engine actually is

An answer engine is any surface that responds to a question with a composed answer rather than a list of links. Today that means ChatGPT, Claude, Perplexity, Gemini, Google AI Mode, Google AI Overviews and Microsoft Copilot. The category also includes older forms, featured snippets and voice assistants, which is why the term AEO predates the current wave. If you want the mechanics of how these systems ingest a page, we covered that in how AI engines index and parse content.

The important consequence is structural. A results page gives ten candidates and lets the buyer choose. An answer gives three to five names and an implied endorsement. There is no page two. Either you are in the answer or you are not in the consideration set, and the buyer has no way of knowing you were omitted.

AEO, GEO, LLM SEO, AI SEO: are these different things?

Mostly no, and the industry has spent a year manufacturing distinctions that do not survive contact with the work.

If a vendor's pitch depends on AEO and GEO being fundamentally different disciplines, that is a marketing position rather than a technical one. Whether the whole category is worth taking seriously at all is a fair question, and we argued it out in is AEO and GEO snake oil.

AEO versus SEO, measured rather than asserted

The claim that SEO does not automatically deliver AI visibility gets made constantly and evidenced rarely. Here is ours.

We asked AI engines who the best managed IT providers were across 15 regions in the US and Canada, resolved every company named to a real website, and checked each against that region's own Clutch directory and against Google results for the same question. The engines named 566 companies across 743 records. The directories held 228 firms between them. The engines named 89 and passed over the other 139. Full method and per-region data is on the research report.

Directory presence is a weak proxy for prominence, so we added a harder control: the 2026 MSP 501, an independently ranked list of the industry's leading firms. Of its 88 ranked firms in those same regions, 66 were named by no engine at all, and zero were named by Claude. These are firms that submitted revenue data to be ranked among the best in their industry, and two thirds of them are absent from the answer a buyer now receives.

The position we hold and have not changed: SEO is the floor and AEO is a distinct layer you can win even when your Google rank does not move. If you are already good at SEO, a fair amount transfers and a critical amount does not, and if you are worried about the traffic side of this shift we covered that in will AI search kill your SEO traffic.

How engines actually pick who to name

There are two mechanisms behind the answers, they behave completely differently, and almost nobody separates them. Which one you are facing changes what you should do this quarter. The underlying idea is old: Lewis et al. showed in 2020 that letting a model fetch documents before answering produces more factual, more specific output than generating from parameters alone.

The corroboration rates follow the same split, inverted. When we checked whether the companies each engine named could be independently verified, Perplexity led at 39 percent, or 101 of the 257 names it produced. Gemini managed 18 percent, ChatGPT 7 percent, and Claude just 2 of its 129 names. Six companies named in the study have no trace of existing anywhere, and all six were Claude's. The full argument is in the caveat that tells you an engine is not looking you up, and the failure mode it produces is covered in when AI gets your company wrong.

One caveat on our own method, because it matters for how you read anyone's engine coverage claims including ours: we collect by calling models through their APIs, which is how any monitoring tool collects at volume, and in that mode two of the six have no web access attached. The consumer apps often behave differently. We wrote that up separately in why the API and the chat window give different answers.

The engines do not agree with each other

This is the finding that breaks most measurement approaches. We asked four answering engines to name the best personal injury firms across five US metros and collected 136 firm-and-market entries. Exactly one engine accounted for 85 percent of them, and no firm at all made every list. On broader category-level questions put to the same engines on the same day, two firms were named unanimously.

Agreement is a function of how specific the question is. Buying questions are specific, so at the bottom of the funnel you are running four separate races, not one. That is why a blended visibility score is a trap and why which engines you track is a real decision rather than a feature-count. Full data in every engine gave a different shortlist.

Does any of this traffic convert yet?

The best public dataset on this is Orbit Media's analysis of 97 B2B GA4 accounts covering 28.9 million sessions from July 2025 to June 2026, with every site required to have at least 100 AI-referred sessions.

Read both halves. The channel is tiny and the visitors are the best you will get, which is what you would expect from someone who arrives having already been given an assessment of their problem and a recommendation. The authors are explicit that the real share is higher than 0.5 percent, because Google AI Mode and AI Overviews are recorded as organic search in GA4 and app traffic lands in direct. We hit the same wall measuring our own: see how to track AI referral traffic in GA4 and cited by AI but getting no clicks.

What actually moves AEO, in order

Deeper on the individual levers: the six types of AI citation, whether backlinks matter, reviews and AI visibility, FAQ schema, llms.txt and robots.txt, and the AI crawlers themselves.

Where the published numbers do not hold up

A definitive article has to check the numbers it repeats, including the popular ones. The clearest failure in this field is Reddit.

For a year this category has repeated one figure: Reddit is around 12 percent of ChatGPT citations. It is the basis for a great deal of advice telling B2B firms to go and post on Reddit. We went looking for the study behind it and could not find one, so in August 2026 we removed it from every page on this site rather than keep passing it on. If you are quoting it, ask whoever you got it from for the method.

Then in July 2026 Dan Petrovic published a citation-mining study using OpenAI's grounding metadata, which uniquely exposes both selected and unselected sources. Across 27,351 probes producing 103,974 fan-outs, he measured OpenAI supplying Reddit as a candidate 491,024 times and citing it 3,012 times: a 0.61 percent selection rate and a 99.39 percent rejection rate. Reddit was among the candidate sources in 76 percent of OpenAI's searches and was thrown away almost every time.

Read those two rates together, because the pairing is the finding. Retrieved constantly, chosen almost never. And a 0.61 percent selection rate is not the same claim as "Reddit is 0.61 percent of ChatGPT citations", which is a different denominator again. Mixing denominators is exactly how a number like 12 percent survives unchecked for a year.

What is not in dispute is the per-engine variation, and it is dramatic. In the same study Reddit accounts for about 2 percent of Google's cited sources, second only to YouTube, while across 139,601 Anthropic grounding sources between May and July 2026 Reddit does not appear once. That matches our own finding that Claude behaves nothing like the retrieving engines. The practical reading: Reddit is a Google play far more than a ChatGPT play, and not a Claude play at all. Our own take on community mentions is in Reddit, G2 and third-party mentions.

How to measure AEO without fooling yourself

Sample, do not spot-check. Answers are non-deterministic. Ask the same engine the same question twice and you can get two different shortlists with nothing changed on your side. A single run is a coin flip wearing a percentage sign. See how often to check.

An engine returning nothing is not evidence of absence. Timeouts, refusals and empty responses must be their own state and excluded from the denominator, not scored as "not mentioned". Google AI Overviews in particular answered only around 9 percent of the commercial vendor queries in our data, against roughly 50 percent of questions about a named brand, because Google frequently renders no AI Overview on a commercial query at all.

Keep the questions stable between checks. If the question set regenerates each month, this month and last month are two different tests. We watched two consecutive scans of one brand share zero questions and turn an 8 percent result into a zero, which looked like a collapse and was two unrelated measurements.

Practical guides: how to measure AI visibility, the free way to check, and if you are evaluating a vendor, eleven checks to run on any AI visibility report.

Why this is worth your quarter

G2's 2026 buyer research found 51 percent of B2B buyers now begin vendor research on an AI chatbot, up from 29 percent, that 69 percent changed their vendor choice based on what an AI told them, and that one in three ended up buying from a vendor they had never heard of before the AI named it. Forrester found twice as many buyers name AI as their top research source than name any other source, including your own website.

That one-in-three number is the one to sit with. A third of these deals begin with the buyer having no prior awareness of the company they end up choosing. Awareness did not precede the decision. The answer created it. AEO is the discipline of being the company created into the consideration set at that moment, and on the timeline question the honest answer is weeks on the retrieval engines and unknowable on the rest. In funnel terms an engine now returns a bottom-of-funnel answer to a top-of-funnel question, which is the shift we set out in the TOFU, MOFU, BOFU guide.