Measuring whether AI assistants mention you at all

Two separate questions hide behind "are we visible in AI". The first is whether assistants send you traffic, which is measurable in your existing analytics. The second is whether they mention you when they do not send traffic. For Copilot, Bing now reports that directly; for most other assistants it still has to be observed.

The two platform reports that exist

Both are free and worth connecting. Neither covers every assistant, and neither tells you whether you were described correctly — which is why the rest of this article still applies.

Identifying assistant referrals, which you already have

Visits originating from an assistant's cited link arrive with a referrer identifying the assistant. Most analytics setups bucket them into generic referral traffic where nobody looks, so the data exists and is unread.

1. 01Pull the referrer report for the last ninety days with no minimum-sessions filter, since these volumes start small.
2. 02Identify the assistant and AI-search hostnames in the list and group them into one segment.
3. 03Compare that segment's behaviour against organic search — pages per session, time on page, conversion rate.
4. 04Review it monthly, because the set of hostnames changes as products launch and rename.

Two caveats keep this honest. Some assistants strip or obscure the referrer, so this undercounts by an unknown amount. And a visit that arrives by typed URL after someone read an uncited answer is indistinguishable from any other direct traffic — which is exactly the case this measurement cannot see.

The citation panel: a fixed question list, run on a schedule

For every assistant the platform reports do not cover, a structured manual audit is the only instrument available. Its value comes entirely from being repeatable, which means the questions must be fixed before the first run.

That last row is the methodological error most people make. Asking follow-up questions in the same conversation conditions every answer after the first, so what gets recorded is the effect of the conversation rather than the assistant's unprompted response.

One row per question, per assistant, per run:

date | assistant | question | cited us? | sources cited | answer correct? | notes

The four columns that matter, in order of usefulness:

1. answer correct?   A wrong answer about your own offering is the
                     most actionable finding in the whole exercise.
2. sources cited     Tells you who currently occupies the answer, which
                     is a far more specific competitive picture than a
                     ranking report.
3. cited us?         The headline number, and the noisiest one.
4. notes             Where the real information is: how you were
                     described, what was omitted, what was inferred.

Reading the result

• Cited, described accurately — working; nothing to do but keep the page current
• Cited, described inaccurately — the page is being read and misunderstood; usually the answer is buried, hedged or split across sections
• Not cited, competitor cited — read the cited page and identify what specific, attributable claim it has that yours does not
• Not cited, nobody cited — the answer was composed from general knowledge, which means there is no established source. That is the opportunity
• Answer is wrong about us — highest priority regardless of everything else

When an assistant is wrong about your company

This happens, it costs enquiries, and there is no correction mechanism to appeal to. The available response is to remove the ambiguity the wrong answer was generated from.

1. 01Find where the wrong fact could have come from. A stale directory listing, an old profile, a similarly named company, or a page of your own that implies it.
2. 02State the correct fact plainly on your own site, in one sentence, on a page where it belongs. An inference happens where a statement is absent.
3. 03State it consistently — the same fact in the same form on every page that touches it, and in your structured data.
4. 04Fix the third-party sources you control. Old profiles and directory entries with former names or addresses are frequently the actual origin.
5. 05Re-run the panel next month. This changes over time rather than immediately; a single re-check proves nothing either way.

The pattern behind nearly all of these is the same: the model inferred something because nothing stated it. Ambiguity is filled in, and it is filled in from whatever source was most available.

Reporting this honestly

A small manual sample run monthly is real evidence and weak evidence, and it should be presented as both.

The third row is the tempting one, and it is how a reasonable audit becomes a misleading report. Three to six out of twenty is a real improvement worth showing; expressed as a percentage change it implies a precision the method does not have.

Run the panel, record what happened, fix the wrong facts first and the uncited questions next. That is the whole discipline, and it is considerably more than most organisations currently do.