ResearchMeasurement research

Is your AI visibility actually improving?

Before trusting a rising score, check the questions, the answers, and what changed in the reporting. Here’s how I’d review it with your team.

Here’s where I’d start

A useful report should explain where you appear, what the answer says, and whether the change matters to your buyers. Keep the question set consistent, separate brand from discovery, and put platform data beside your own visits and inquiries. That gives you something you can act on.

  1. 01Ask the right question
  2. 02Read the right measure
  3. 03Explain the change
  4. 04Choose the next action

Decide what you want the number to tell you

You’ve got a higher AI visibility score this month. What improved? Did more buyers have a chance to see your business? Did the answer describe you more accurately? Did you simply track different questions? I’d settle that before calling the month a win.

Google can report a displayed link. A monitoring tool can save what a selected question produced. Website analytics can record a visit, and your CRM can record an inquiry. Each is useful, but each sees a different part of the journey.

Start the review with the business question: are suitable buyers discovering us, getting the right information, and taking a useful next step? Then choose the evidence that can answer it. You’ll get a clearer report than putting every available number on one chart.

Google Search Console: Generative AI performance report (Search)

Check how Google counts an AI impression

Google says its Generative AI performance report became available worldwide on 31 August 2026. It covers link impressions in AI Overviews and AI Mode, excluding Search Labs experiments. It doesn’t count every mention of your name in the answer text.

The default chart counts at property level. Two links from the same site in one feature count as one chart impression. Page tables count at page level, so adding their rows can give a different total. The AI report also overlaps with Web reporting. Adding both totals would count some exposure twice.

Watch the export too. Values displayed as a dash or tilde become zeros in downloaded data. If a spreadsheet says zero, check whether the platform actually measured no impressions or didn’t have a usable value.

I’d put the counting rule beside the chart, especially when a report goes to leadership. You shouldn’t need to be a Search Console specialist to understand what the number represents. Check the original view before you explain a jump or a disappearance.

Google Search Console: Generative AI performance report (Search)

Treat Bing’s topics as a research aid

Bing launched Topics Portfolio in beta on 13 October 2026. Topic Size approximates citations across sites for a topic. Queries counts distinct grounding queries, while Sites Cited estimates the sites participating. Opportunity volumes are given as ranges.

Grounding queries are the searches used to find supporting material. They aren’t a count of people asking the same buyer question. A large topic in this report doesn’t automatically mean a large audience ready to buy from you.

Some topics are left out when site presence or wider activity is too low. So a missing local topic doesn’t prove there’s no interest in it. I’d check its relevance with the people who know the market before removing it from the strategy.

Use the topic view to find pages and questions worth investigating. Then check the actual answers, your commercial fit, and other evidence of demand. It gives you useful context without replacing those checks.

Bing: Introducing Topics Portfolio (Beta) in AI Performance

Open the answer behind a citation

Bing’s original AI Performance announcement describes citation counts, pages, query samples, and trends. It explicitly separates those measures from ranking, authority, or placement in an answer. That’s an important distinction when someone asks whether the brand is “winning.”

Your page might be cited for a useful explanation while the answer recommends someone else. A care-funding page could help explain the rules without putting the provider on the family’s shortlist. The source did a job; it may not be the job you were measuring.

The same applies to software documentation and retail buying guides. I’d check whether the business is described, compared, recommended, or criticised, and whether the source supports what’s said. A single citation count hides those differences.

Keep a few saved answers in the leadership report. Show one useful improvement and one unresolved gap, with the question and source. That makes the numbers much easier to discuss and turns the review into decisions about the work.

Bing: Introducing AI Performance in Bing Webmaster Tools—Public Preview

Does the source back up what AI said?

Open the link and check the exact claim before you repeat it.

What does it say?

Write down the statement you’re checking. Visits, inquiries, pipeline, revenue, and profit mean different things.

Open the community facts workbook

Split questions about your name from discovery

Our anonymised Canadian senior living sample has 17 brand mentions across 100 generated ChatGPT questions. At first glance, that is a 17% mention rate. The split is more useful: 13 out of 13 questions naming the operator mention it, compared with four out of 87 questions without its name.

That makes discovery about 4.6%. Around 76.5% of the observed mentions come from questions that already give the model the operator’s name. We checked the sample on 3 November 2026; the original collection period wasn’t shown. These are answer counts, not a measure of family reach.

If I added more questions with the operator’s name, the overall score could improve even if discovery stayed exactly the same. I’d keep the groups separate and show how the question set changed whenever I introduced new coverage.

Treat the question list as part of the result. Save the wording, topic, market, language, engine, and date. Otherwise the next review may be comparing two different tests.

Anonymized senior living visibility dataset and lesson
Split questions about your name from discoveryTry this
Question groupMentions / questionsWhat this sample suggests
13 / 13Recognition when the name is supplied
4 / 87Limited discovery across these sampled questions
17 / 100A total that conceals the difference

AI knows the name. Does it introduce the business?

SearchableWhere the brand appears
ChatGPT
Canadian retirement operatorReport snapshot
Mention rate17%

17 of 100 questions mention the brand

Here’s the split: 13 of 13 questions with the name mention the operator. Only 4 of the other 87 do.

Why look at them separately?

Ask about your brand and AI may know it well. Ask for suitable options and you might not appear. One combined score can hide that gap.

Which questions bring it up?

TopicQuestionsMentionsRate
23313%
1700%
131184.6%
12216.7%
1200%
800%
7114.3%
600%
200%
About these numbers

100 generated ChatGPT questions. Checked November 3, 2026; collection date not shown. The prompt table labels 13 branded questions (13 mentions) and 87 unbranded questions (4 mentions). Topics and question types are different groupings. Rates rounded. One snapshot, not market share, consumer search volume or care quality.

The suggested next steps are my advice, rather than results measured in this report.

Don’t turn a sample percentage into a traffic forecast

Picture two equal-sized test sets. One mostly asks about your company by name. The other asks people to discover suitable suppliers. You could perform the same within each type and still receive different overall scores because the questions are mixed differently.

Neither test tells you how many real people ask those questions or see the resulting answers. It also doesn’t tell you whether the mention was helpful enough to prompt a click. Those are separate things to observe.

I’d use the sample to test the answer, then use first-party records to follow visits and inquiries. Where the path isn’t visible, say so. You can still learn something useful about discovery without filling the missing steps with assumptions.

A stronger score should give you a better question to investigate: which answers improved, for which buyers, and what happened next? It isn’t a percentage you can apply directly to your market size or website revenue.

Anonymized senior living visibility dataset and lesson

Explain a chart movement before reacting to it

Google recorded a logging error in Generative AI Search impressions for 13–17 August 2026. Its 21 August update says the missing data was restored. The problem affected the reporting, rather than actual search visibility.

If you saw that decline while it was happening, rewriting your content wouldn’t have fixed it. I’d first check reporting notices, recent data completeness, settings, and changes to the tracked questions. Then look at what changed in the answers themselves.

That doesn’t mean every decline is a reporting problem. Competitors, content, the platforms, and buyer questions can all change too. It means the explanation should fit the evidence, rather than whichever marketing activity happened closest to the dip.

The August incident is resolved. Keep it as an example of a check worth making, rather than an excuse to dismiss a later fall. If the same answer gaps keep appearing across repeated checks, they deserve attention.

Google Search Console: Data anomalies

Connect the report to a useful business outcome

Searchable separates brand mentions from linked sources and lets teams inspect the answer evidence. I’d use that to understand where you’re missing and what the system says about you. The next question is whether the change helped someone make a better decision.

In senior living, that means looking at suitable inquiries, tours, and move-ins separately. A prominent recommendation for care you don’t offer can create poor inquiries. A corrected service explanation can save the team time even if lead volume doesn’t rise.

Agree what matters with sales or admissions before the review. That might be clearer expectations, discovery in a priority location, or better inquiry fit. Then show the evidence for that outcome instead of relying on the most flattering chart.

Keep credited revenue separate from revenue you believe the work added. A CRM source label helps describe the journey; it doesn’t prove the customer would never have arrived otherwise. The size of the claim should match the comparison you can support.

Searchable: AI Mentions & Citations Analytics

What happens after the AI answer?

Pick a stage. Check what you can measure there, then follow the next step in your records.

Does the right community appear, and are its details right? That shows what the answer said. It doesn’t tell you how many families saw it.

Try the revenue calculator

End the review with something to do

I’d keep the report to a few connected questions: what did we track, which answers changed, which pages gained exposure, and what happened to the inquiries we can follow? Keep the groups and dates visible so someone can compare the next month fairly.

Then choose the next action. Fix a recurring wrong claim. Investigate why a competitor appears for a relevant comparison. Check whether a new page is being found. Each should have a named owner and a clear question for the next review.

This guide uses Google and Bing documentation plus our supplied answer sample, checked on 3 November 2026. The equal-sized test sets are an illustration, not a campaign result. The source notes keep the platform definitions and limits available.

I can help you set up that review in Searchable and make the findings useful to the team. Next, take a closer look at citations. Once you know where a source appears, you still need to check whether it supports what the answer says.

Searchable: AI Mentions & Citations Analytics

Sources and details

Sources reviewed: November 3, 2026. I work at Searchable. The examples and recommendations are mine; reported results are credited to their sources.

1
Google Search Console: Generative AI performance report (Search)Official documentation reviewed November 3, 2026; global rollout dated August 31, 2026. Covers AI-feature link impressions, aggregation, overlapping Web data, and export treatment of unavailable values.
2
Bing: Introducing Topics Portfolio (Beta) in AI PerformancePublished October 13, 2026. Approximate citation-based topic measures and visibility thresholds; not counts of end users or consumer prompt volume.
3
Bing: Introducing AI Performance in Bing Webmaster Tools—Public PreviewPublished February 10, 2026. Official scope and metric definitions for supported Microsoft AI experiences. Citations do not measure rank or placement.
4
Anonymized senior living visibility dataset and lessonSupplied report recounted November 3, 2026. One Canadian operator, 100 generated ChatGPT questions, 17 mentions. Question composition calculations are ours. Collection date and window unavailable; not consumer research or market share.
5
Google Search Console: Data anomaliesReviewed November 3, 2026. Generative AI Search impression logging issue dated August 13–17, 2026; August 21 update confirms restoration. No effect on actual visibility reported.
6
Searchable: AI Mentions & Citations AnalyticsOfficial feature description reviewed November 3, 2026. Supports examining mentions and sources separately. Commercial and causal interpretation is our analysis; I work at Searchable.
Searchable
Daavid ChristaYour partner at Searchable

Let’s work out where to start.

I’m Daavid, and I work at Searchable. Bring your markets, care services, and the questions families ask. We’ll build a useful question set, check the answers, and agree what to improve first.

See where you stand

What you get
A starting view of how AI mentions your business, with the market, model, and dates clearly shown.
Who takes part
You and Daavid
How Searchable helps
Compare saved answers, competitors, and citations.
Talk it through with me45-minute conversation

We’ll agree what your team needs, what’s included, and when to review it.