AI brand monitoring: what the platforms actually report

Classic media monitoring counts mentions. AI answer engines mostly do not expose mentions — they expose citations, which is a narrower thing: a link to your site shown as a source. Before buying a tool or setting a target, it is worth knowing exactly which numbers the platforms themselves publish, because that is the only first-party evidence you will get.

Two platforms publish anything at all today. They disagree about what is worth counting, and the gap between them is where most confusion about "AI share of voice" comes from.

What is actually established

Everything below is from the platforms' own documentation, not from vendor benchmarks.

Google reports impressions only. The Generative AI performance report counts "how many times links to your site were shown to a user in a generative AI feature on Google Search," across AI Overviews and AI Mode. It breaks down by pages, countries, dates and devices.
Google gives you no query data there. The report has no clicks, no CTR, no position and no query dimension. AI-feature traffic is also folded into the ordinary Performance report under search type "Web" rather than split out.
Not every property has it. Google states the report is rolling out over time, so an empty report is not evidence of zero visibility.
Microsoft reports citations, and one share metric. Bing Webmaster Tools' AI Performance covers Copilot, AI-generated summaries in Bing and select partner integrations, with Total Citations, Average Cited Pages and page-level citation activity.
"Grounding queries" are not user prompts. Microsoft defines them as the phrases the AI used when retrieving content — the retrieval layer, not what a person typed. Treating them as keyword data is a category error.
Citation Share is the closest thing to share of voice — "the percentage of citations attributed to your site out of all citations shown across all sites for that same grounding query." Microsoft is explicit that it is "designed as an observational metric – not a ranking system or a competitive scoreboard. It does not expose competitor domains, represent traffic share, or assign quality scores."
Opting out is scoped and reversible. Google's Search generative AI control governs appearance in AI Overviews, AI Mode and generative AI in Discover; opting out forfeits impressions and traffic from those surfaces, is not a ranking signal elsewhere, and does not affect AI training.

How to run it

The aim is a baseline you can defend in a review, not a dashboard number nobody can source.

Turn on both reports before you need them. Neither backfills, so the first month you look is the earliest month you will ever have.
Record Google impressions by page, not by brand. That is the only dimension Google gives you, so build the reporting around it rather than around queries you cannot see.
Record Bing Total Citations and, where available, Citation Share by grounding query. Label these clearly as citations, never as mentions.
Keep the two sources separate in every report. They count different events on different surfaces; a combined "AI visibility score" destroys the only property that makes them useful — knowing what was measured.
For anything the platforms do not report — whether a model names your brand in prose, and how it characterises you — run your own prompt sample on a fixed prompt set, a fixed date and a stated model version, and label it as sampling.
Re-run that sample on the same prompts each month. A moving prompt set produces a moving number that means nothing.

How to measure it

Pick metrics you can point at a source for.

AI-feature impressions per page, per month, from Search Console — the trend matters more than the absolute number.
Total Citations and Average Cited Pages from Bing Webmaster Tools.
Citation Share per grounding query, read as an observation about that query, not as a competitive standing.
Your own prompt-sample results, versioned by model and date, reported as a sample with its size stated.
Assisted outcomes: branded search volume and direct traffic, which are the places where answer-engine exposure without a click tends to surface.

Where teams get this wrong

Reporting citations as "mentions." Neither platform reports whether a model said your name in the answer text; both report link citations.
Buying a "competitor share of voice in LLM answers" number and treating it as platform data. Microsoft explicitly does not expose competitor domains, and Google exposes nothing at query level — so a named-competitor figure comes from a vendor's own prompt sampling, with that vendor's prompt set and sampling error.
Reading an empty Google report as zero visibility while the feature is still rolling out to properties.
Treating grounding queries as search keywords and briefing content against them.
Blocking crawlers to "protect" the brand without checking which surface each token governs — the controls are not interchangeable, and one of them costs you search visibility while another does not.

Sources

AI brand monitoring: common questions

Can I see which prompts caused my brand to appear?

Not from Google — the generative AI report has no query dimension. Bing publishes grounding queries, but those are the phrases the AI used to retrieve content, not what the user typed. Anything prompt-level beyond that comes from your own sampling.

Is there an official share-of-voice metric?

Microsoft's Citation Share is the only one, and it is share of citations for a grounding query, not share of traffic or of brand mentions. Microsoft states it is observational and does not expose competitor domains.

Do OpenAI or Perplexity give publishers visibility analytics?

We found no first-party visibility reporting from either that we could verify on their own documentation. Treat any figure about ChatGPT or Perplexity visibility as third-party sampling unless the platform documents it.

Should we opt out of AI features to protect the brand?

Google's control removes your content from AI Overviews, AI Mode and generative AI in Discover, and you lose the impressions and traffic from those surfaces with it. It is not a ranking signal elsewhere and it does not affect AI training, so it is a distribution decision, not a safety one.