TL;DR
- AI Overviews visibility splits into three states: presence of an AI Overview at all, a citation linking your domain, and a mention of your brand with no link. A mention produces no data in any Google report.
- Search Console's main Performance report folds AI Overview clicks into Web totals with no filter; its separate generative AI report shows impressions and pages but no clicks or queries. Position for a cited page reflects the AI Overview block, not your place inside it.
- Track a fixed panel of 30 to 50 queries, frozen for a quarter, recording presence, your status, every cited domain, the date and the location.
- Competitor share of voice needs no competitor analytics: count cited domains across the same panel over a quarter.
- Report monthly, not weekly. Model updates move every number on their own, so act on a trend across at least a quarter.
To track AI Overviews visibility you need three separate things: a fixed set of queries you check on a schedule, a record of whether your brand was cited, mentioned or absent in each answer, and Search Console data to tell you what that visibility did to clicks. No single tool gives you all three, and Search Console's own generative AI reporting does not separate AI Overviews from ordinary results in the way most people assume it does. This page sets out what can actually be measured, what cannot, and a routine that survives contact with a real reporting cycle.
What AI Overviews visibility means, precisely
The phrase covers three different states, and conflating them is the first thing that breaks a tracking programme.
| State | What it means |
|---|---|
| Presence | An AI Overview appeared at all for that query, regardless of who is in it |
| Citation | Your domain appears as a linked source under or inside the answer |
| Mention | Your brand name appears in the answer text without a link to your site |
Presence is a property of the query, not of you. Citation is a link you can measure in Search Console indirectly. A mention is neither: it generates no click, no impression attributable to your site, and no entry in any Google report. If your brand is named in an answer and a competitor is cited, you will see nothing in your own data at all. That gap is the single hardest part of this measurement problem, and it is the reason a manual or tool-based check of the answers themselves is not optional.
What Search Console shows, and what it does not
Search Console gives you two things, and neither is the filter people expect. In the main Performance report, clicks and impressions from an AI Overview link are folded into the same Web search type totals as ordinary blue-link results, with no dimension that says this row is an AI Overview. Separately, an expandable generative AI performance report sits under the Performance report. It counts impressions of your URLs in generative AI features on Search and Discover, split by Pages, Countries, Devices for Search only, and Dates down to the hour. It carries no click data and no query data, it does not distinguish AI Overviews from AI Mode or from Discover's AI features, and it only appears once a site clears an impressions threshold Google has not published. So it can tell you which pages Google's AI surfaces are drawing on and how often; it cannot tell you what was asked or whether anyone clicked.
Position reporting adds a second complication. Where your page is linked from an AI Overview, the position recorded reflects the position of the AI Overview block itself rather than your link's place within it. So a page cited third in an AI Overview at the top of the page can report a strong average position while receiving very few clicks, which looks like a click-through-rate problem and is really a reporting artefact.
What Search Console is genuinely good for is the second-order signal. If a query cluster holds its impressions and average position while its clicks fall over a period of weeks, that pattern is consistent with an AI Overview absorbing the answer. It is not proof, because seasonality, a layout change or a competitor's new page produce similar shapes, but it tells you which clusters deserve manual checking.
Building a query set worth tracking
The instinct is to track everything. That produces a number nobody trusts, because the set drifts. Build a fixed panel instead.
- Start from Search Console. Export queries with meaningful impressions where your pages already rank on page one, since those are the queries where losing the click actually costs you something.
- Add the questions your sales team is asked. Those are the prompts a buyer types into an AI answer engine, and they are frequently absent from keyword tools entirely.
- Keep it small enough to check properly. Thirty to fifty queries checked consistently beats five hundred checked once.
- Freeze the set for a quarter. Every addition or removal breaks the comparison with last month, so batch changes to a quarterly review.
- Record the query in the language your customers actually use. For a Thai business that usually means running two panels, Thai and English, because the answers differ and so do the sources cited.
What to record for each check
Consistency matters more than sophistication. Five fields per query per check are enough.
| Field | Why it is there |
|---|---|
| AI Overview present, yes or no | Presence rate for your category is the denominator for everything else |
| Your status: cited, mentioned, absent | The three states above, recorded separately, never merged |
| Which domains were cited | Competitor share of voice comes from counting these over time |
| Date and location of the check | Answers vary by location and change without notice, so an undated reading is worthless |
| A copy of the answer text | The only way to see later what changed, and the only defence against a disputed claim |
Share of voice then falls out of the citation column: for the queries where an AI Overview appeared, what percentage cited you, and what percentage cited each competitor. That is a defensible number because you can show the working, which matters when someone asks how it was calculated.
Manual checking versus tools
Manual checking is free, gives you the actual answer text, and lets you see the whole competitive set. It is also slow, and it is not reproducible: two people checking the same query minutes apart can get different answers, because the models are not deterministic and results are personalised and localised.
Tools automate the volume and give you a trend line, which is what a monthly report needs. Their limits are worth stating plainly. They sample, so their presence rate is an estimate. They check from their own locations and profiles, which may not match your customers'. Different tools disagree with each other on the same query, because they are measuring slightly different things. And a model upgrade can move every number overnight without anything about your site changing.
The practical answer is both: a tool for the trend across the whole panel, and a manual spot check of ten queries a month to keep the tool honest and to read what the answers actually say about your brand.
Measuring competitors
Competitive share of voice in AI Overviews is easier to measure than your own click impact, because it needs no access to anyone's analytics. You are reading a public answer and counting citations.
Run the same fixed panel, record every cited domain rather than only your own, and count over a quarter. Two patterns usually emerge. One is that a handful of non-competitor domains, typically reference sites, marketplaces and forums, take a large share of citations in most categories. The other is that the competitor set in AI answers is often not the competitor set from organic rankings, because the answer engine is assembling from sources rather than ranking pages.
Both findings change what you do next more than your own citation rate does. If a marketplace listing is cited where your product page is not, the work is on the marketplace listing.
Connecting visibility to traffic and leads
The question every report eventually reaches is whether any of this produced revenue, and the honest answer is that the attribution chain is broken by design. A person who reads an answer, does not click, and comes back two days later through a branded search is invisible to the first touch.
Three proxies are worth tracking together, none of which is conclusive alone. First, branded search volume in Search Console: a rise with no campaign behind it is consistent with people learning your name somewhere that does not send a click. Second, direct and organic landing-page traffic on the specific pages that are being cited. Third, the question your sales or enquiry team can simply ask, which is where the enquirer heard about you. In Thailand, where most enquiries arrive through a LINE Official Account rather than a form, that last one is often the most reliable of the three, because a human is already in the conversation.
A reporting cadence that holds up
| Frequency | What to do |
|---|---|
| Weekly | Nothing. The answers are noisy enough that weekly readings mostly measure variance |
| Monthly | Run the full panel through your tool, spot check ten queries manually, record citations and competitors |
| Monthly | Pull the same query cluster in Search Console and note impressions, clicks and position side by side |
| Quarterly | Review the panel itself, add and remove queries in one batch, and re-baseline |
Report the three states separately, always show the denominator, and date every figure. A report that says citation share rose from 12% to 18% without saying across how many queries, checked where, and on what date is not a measurement.
Common mistakes and what they cost
- Treating an AI Overview appearing as a loss. Presence is a property of the query. If your page is the cited source, an AI Overview can be a visibility gain.
- Comparing months with different query panels, which makes any change unreadable.
- Reading Search Console position as your rank inside the AI Overview when it reflects the position of the block.
- Counting mentions and citations as one number, which hides the fact that mentions produce no traffic at all.
- Checking from the wrong location. For a Bangkok business, a reading taken from a US data centre is measuring somebody else's answer.
- Reacting to a single month. Model updates move the numbers on their own, so act on a trend across at least a quarter.
What this means for Thai businesses
Thai-language answers behave differently from English ones and have to be tracked separately. The source pool for Thai queries is smaller, which cuts two ways: it is easier for a well-structured Thai page to become a cited source, and a single dominant source can hold a large share of citations in a whole category. Running an English-only panel for a Thai business measures the wrong market.
Location settings matter more here than in a large market. Answers for a query about a service in Bangkok differ from the same query checked without a location, and for multi-branch businesses the difference between Bangkok and a provincial city is often larger than the difference between two competitors.
The work that changes the numbers sits with content structure and citability rather than with tracking, so measurement should be scoped as the input to that work. Our generative engine optimisation team handles the citability side, the technical foundations come out of an SEO audit, and the traffic and conversion measurement behind it depends on a clean GA4 setup. Brands whose main organic programme is also affected should read the two together with their AI SEO plan.
Frequently asked questions
Can I see AI Overviews data in Search Console?
Partly. The main Performance report folds AI Overview clicks and impressions into Web totals with no filter to isolate them. The separate generative AI performance report, an expandable tab under Performance, shows impressions and the pages involved by country, device and date, but has no click data and no query data and does not split AI Overviews from AI Mode or Discover. For traffic impact the usable signal is still indirect: stable impressions and position with falling clicks across a query cluster is the pattern worth investigating.
Why does my position look good but my clicks are low?
Where a page is linked from an AI Overview, the position recorded reflects the position of the AI Overview block rather than your link's place inside it. A page cited within a top-of-page AI Overview can therefore report a strong average position while receiving few clicks.
How many queries should I track?
Thirty to fifty, checked consistently, is more useful than several hundred checked once. Choose queries with meaningful impressions where you already rank on page one, add the questions your sales team actually receives, and freeze the panel for a quarter so month-to-month comparisons hold.
Do I need a paid tool to track AI Overviews visibility?
No, but a tool makes the trend line affordable. Manual checking gives you the real answer text and the full competitive set at the cost of time and reproducibility. The workable combination is a tool for the whole panel plus about ten manual spot checks a month.
How do I measure competitor visibility in AI Overviews?
Record every cited domain, rather than only your own, across the same fixed query panel, then count citations by domain over a quarter. It needs no access to competitor analytics because you are reading a public answer, and it usually shows that reference sites and marketplaces hold more citation share than direct competitors do.
Where to start this week
Pick thirty queries from Search Console where you rank on page one and impressions are worth having. Check each one manually, from your customers' location, and record presence, your status, and the cited domains in a dated sheet. That single sheet is a baseline, and a baseline is the thing most AI visibility programmes never actually create. Everything after it is repetition.
