Are you cited in AI answers? How to check, honestly

A growing share of buyers ask an AI assistant before they ask a search engine. Whether your brand appears in those answers is a legitimate competitive question.

It is also a question that produces bad evidence very easily, because AI answers are non-deterministic and most people check them once.

Why one check is not a finding

Ask the same question twice and you can get different answers, with different brands cited. Change the phrasing slightly and the response can change substantially. The same question asked at different times can produce different results as underlying indexes and models update.

So “I asked ChatGPT and we did not come up” is an anecdote. It might be true in general. From one observation you cannot tell.

The same applies in reverse, and this is the more expensive mistake: one appearance is not evidence that you are reliably cited. Somebody checks once, sees the brand, and reports that the company is well represented in AI answers. That claim will not survive the next check.

A protocol that produces something defensible

Four decisions, each of which turns a casual check into a repeatable measurement.

Fix the questions in advance

Write out the questions before you run anything, and write them the way a buyer would type them. Not your brand name, which tests nothing except whether the model has heard of you.

Category questions:

• “Who are the best commercial roofing contractors in the Midwest”
• “What should I look for in an industrial maintenance provider”
• “How do I evaluate a compliance inspection vendor”

Six is a reasonable set. Enough to cover different framings, few enough to actually repeat.

Fix them in a document. Reuse the exact same set next quarter. A moving question set makes quarter-over-quarter comparison meaningless.

Repeat each question

Three times each, minimum. Fresh session each time, no conversation history carried over.

This is the step that converts an anecdote into a measurement. With three repetitions per question you can distinguish between never appears, sometimes appears, and reliably appears, which are three genuinely different competitive positions.

Record the date and the surface

Google AI Overviews and ChatGPT are different systems with different retrieval behaviour. Perplexity is different again. A finding about one is not a finding about all of them.

Record which surface, which day. AI answer behaviour changes on timescales of weeks, so an undated finding is close to worthless six months later.

State the finding with its bounds

The claim that the data supports looks like this:

Six category questions were tested three times each against Google AI Overviews and ChatGPT on 2026-06-29. The brand appeared in none of the eighteen responses. Two competitors appeared in more than half.

The claim it does not support:

The brand does not appear in AI answers.

The first is checkable, bounded, and repeatable. The second is a general property nobody can verify.

This distinction matters more than it might seem. Absence in eighteen sampled responses supports absence for those questions on that date. It does not support absence as a general fact about the world, and reports that make the second claim from the first are over-reaching in a way that undermines everything around them.

What to do with the result

The comparative reading is the useful one. Not whether you appear, but whether your competitors do.

Nobody in the category appears. The models are answering from general sources, aggregators or directories rather than from vendor sites. That is a flat finding, and like other flat findings it is an opportunity: the category is unclaimed in this surface.

Some competitors appear consistently and you do not. Worth understanding what those competitors have that you do not. In practice it is usually third-party citation: they are referenced by trade publications, quoted in industry coverage, or they publish original research the category cites. Being talked about by others is doing more work here than anything on their own site.

You appear inconsistently. Better than absence, and fragile. Usually means you are a plausible answer without being a well-established one.

The uncomfortable part about optimisation advice

There is a large volume of advice circulating about how to get cited by AI systems. Some of it is sound and derived from how retrieval works. A good deal of it is confident guesswork.

The honest position: the tactics with the best evidence behind them are the unglamorous ones. Clear, well-structured content that directly answers the question a person asked. Being cited by other sites, because third-party corroboration is doing real work in retrieval. Publishing something original enough that other people reference it.

That is uncomfortably similar to what has always worked, which is why it gets less attention than newer-sounding advice.

For a worked example of applying evidence to one of the newer tactics specifically, see the llms.txt question.

Where this fits in an audit

Whether a brand is cited in AI answers is one input to the Inbound and Content lens, and it is deliberately weighted as one input among several rather than treated as a headline.

It carries medium confidence rather than high, with the caveat attached to the record, because the measurement is inherently variable and pretending otherwise would be exactly the kind of over-claiming the rest of the method exists to prevent.

Every claim in a report carries its confidence and its caveat, on the record itself, where somebody deciding what to do with it will see them. That applies to the strong findings and to this one, which sits in the middle by construction.