Measuring whether AI engines cite you is harder than fixing why they do not

Every B2B marketing team has now run the same experiment. Somebody opens an assistant, types the question a buyer would ask, sees a competitor named and their own product absent, and forwards the screenshot to the channel. It is a useful moment of alarm and it is not a measurement.

The reason it is not a measurement is that answer engines are not deterministic. Ask the same question twice and the sources cited can differ. Ask from a different account, a different country or in a different language, and they will differ more. A single screenshot tells you what happened once.

What a real measurement needs

Treat it like polling rather than like rank checking. You are estimating the probability that your brand appears for a class of question, which means samples, repetition and a stable method you do not quietly change between reporting periods.

• A fixed prompt set, written the way buyers actually ask rather than the way you would like them to, covering comparison, alternatives, pricing and integration questions

• Repetition, since the same prompt run twenty times over a fortnight gives you a rate rather than an anecdote

• Coverage of each engine separately, because they retrieve differently and a strong position in one says almost nothing about another

• Locale variation, since the same question from another country frequently returns a different set of sources

• Citation logging rather than mention logging, because being named in prose and being cited with a link are different outcomes with different value

• A frozen definition of what counts as a win, agreed before you start reporting to anyone

The infrastructure problem nobody budgets for

Running a few thousand prompts a week from one office address does not work for long. Rate limits arrive first, then challenge pages, and the run that was supposed to produce a clean weekly number produces a partial one instead. Worse, a partial run looks like a change in visibility rather than a change in your own collection, and teams have redirected content strategy on the back of exactly that mistake.

The usual fix is to spread the work across many addresses and to run from the countries you actually sell into. Since this is public page fetching at volume rather than anything requiring a consumer connection, rotating datacenter proxies are the cheap and fast option, and the rotation is what keeps a long run from tripping a single address limit. Pace the requests anyway, cache every response, and never re-fetch something you already have, because the cache is also what lets you re-parse a month of data when you realise your extraction logic was wrong.

Once you can measure, the fix is editorial

This is the part the tooling conversation obscures. Engines cite pages that are easy to extract a defensible claim from, and most B2B pages are not. For the broader playbook on getting a company visible across both traditional and AI search once you can actually measure the gap, see how to make your company more visible.

• State the claim in the first two sentences of the section rather than building to it

• Attach numbers to a date and a source, because an undated statistic is unusable to a system that has to attribute it

• Define the entity plainly somewhere on the page, in the form a machine can lift as a definition

• Answer the comparison question directly, including where you are not the right choice, since pages that name their limits get cited by systems asked for balance

• Keep one claim per heading, so retrieval can take a section without taking the whole page

• Publish the specifics competitors leave vague, such as pricing structure, deployment requirements and integration lists

Two ways teams mislead themselves

The first is confusing visibility with attribution. Being cited is not a lead, and a rising citation rate that coincides with a flat pipeline is a signal to look at the question set rather than to celebrate. Instrument the pages that get cited so you know what happens after the click.

The second is running the measurement from a machine that is logged into your own accounts, which quietly personalises what you see. If your measurement can be influenced by whose browser it ran in, it is not a measurement of the market.

FAQ

How many times should I run each prompt?

Enough to get a rate rather than a result. Twenty runs spread over two weeks per engine is a reasonable floor for a prompt you plan to report on.

Do results really change by country?

Frequently, and by language more so. If you sell into several markets, measuring only from your own understates and overstates different things at once.

Is being mentioned as good as being cited?

No. A citation carries a link and attribution, which is what moves traffic and what survives being summarised again downstream.

What is the fastest fix once I can see the gap?

Rewrite the page so each section answers one question with a dated, sourceable claim in the opening sentences. Structure beats volume here.