AI Search Statistics That Survive Checking

Updated September 2026 · Written and maintained by the Progression Agency strategy team

Most AI search statistics are quoted without a date, a denominator or a link to whoever measured them first. We traced them. Five hold up, three are unverifiable company disclosures, and three of the most repeated could not be traced at all.

On this page · 22 sections

Which AI search statistics actually survive checking?
The statistics that hold up
Why does every source give a different AI Overview trigger rate?
How much does the trigger rate vary by industry?
What about the enormous user numbers — 2.5 billion, 1 billion, 800 million?
Which widely-quoted statistics could not be traced at all?
What has nobody measured yet?
How should you read any AI search statistic?
Why we published this instead of another statistics listicle
Is AI search actually replacing Google?
What should you actually do with these numbers?
How do vendor studies and academic studies differ here?
What changed between 2025 and 2026?
Who measured the click-through impact, and can it be trusted?
What is the single most useful number on this page?
Which figures are safe to put in a board deck?
Are the AI Overview trigger rates rising over time?
How many AI search statistics did you check?
What would change our conclusions?
Can we use these statistics in our own marketing?
Why does the wording of a statistic matter so much?
More from Progression Agency

The short answerFive AI search statistics survive tracing with a stated sample and method: Pew’s click-through study, Ahrefs’ and Conductor’s AI Overview trigger rates, Conductor’s AI referral share, and Conductor’s industry breakdown. Alphabet’s user figures are real disclosures with no methodology attached. The ‘800 million weekly ChatGPT users’ figure is not published in any OpenAI dataset — it comes from a keynote. And nobody has measured what share of AI citations come from which kind of source.

Research run 15-16 September 2026 across 107 verification passes against primary sources. Every figure is attributed and dated. Progression Agency sells answer engine optimisation services and therefore benefits if you believe AI search matters — which is why each number is linked to its publisher so you can check it.

Which widely-quoted statistics survive tracing — Good = traced to a primary publisher with a stated method. Part = first-party disclosure with no method. Bad = could not be traced to a primary measurement.

Which AI search statistics actually survive checking?

Five, with a stated method and a traceable publisher: Pew’s click-through study, Ahrefs’ and Conductor’s AI Overview trigger rates, Conductor’s AI referral share, and Conductor’s industry breakdown. Three more are real but are first-party disclosures with no methodology attached. Three of the most-repeated figures in this field could not be traced to a primary measurement at all.

We checked because the numbers in circulation disagree with each other and almost none of them carry a date, a sample size or a denominator. The research ran across 107 verification passes against primary sources, and where a figure could not be re-confirmed we went back to the original publisher’s page and read it directly rather than repeating it.

What follows is organised by how much weight each statistic can carry, not by how often it is quoted. Several of the most-quoted figures are in the weakest category.

Pew Research Center — Verified. 68,879 searches, 900 metered US adults.

Ahrefs — Verified. 20.5% of 146,122,391 desktop SERPs.

Conductor — Verified. 25.11% of 21.9M US searches.

Conductor referrals — Verified. 1.08% of 3.3B sessions, 1,215 domains.

Conductor by industry — Verified. Health Care 48.75% vs Real Estate 4.48%.

seoClarity — Corroborating. Same direction, wider spread.

The statistics that hold up

Each of these has a named publisher, a stated sample, a date, and a method you can go and read.

Verified AI search statistics, with sample and method
Figure	Publisher	Sample and method	When	Commercial interest
Users clicked a search result in 8% of visits with an AI summary, versus 15% without	Pew Research Center	68,879 unique Google searches from 900 US adults on Ipsos’s KnowledgePanel Digital, passively metered browsing	March 2025, published 22 July 2025	None — nonpartisan nonprofit
Users clicked a link inside the AI summary in 1% of visits	Pew Research Center	Same dataset	March 2025	None
18% of searches in the study produced an AI summary	Pew Research Center	Same dataset — real user searches, not a keyword index	March 2025	None
AI Overviews appeared on 20.5% of SERPs (29,991,998 of 146,122,391)	Ahrefs	146.1M desktop SERPs from Ahrefs’ own keyword index	September 2025	Sells SEO software
AI Overviews triggered on 25.11% of searches (5.5M of 21.9M)	Conductor	21.9M unique US Google searches from Conductor’s enterprise keyword index	15 Sept – 12 Oct 2025	Sells AEO software
AI referral traffic was 1.08% of all website traffic (35.7M of 3.3B sessions)	Conductor	1,215 enterprise customer domains	May – Sept 2025	Sells AEO software
Health Care 48.75% versus Real Estate 4.48% trigger rate	Conductor	10 GICS-aligned industry categories	Sept – Oct 2025	Sells AEO software

Pew Research Center: clicks with and without an AI summary — The only click-through measurement in this literature from an institution with no commercial interest in the answer. Published 22 July 2025.

Why the Pew study is the one to cite

It is the only click-through measurement in this literature produced by an institution with no commercial interest in the result. Every other CTR figure comes from a company that sells SEO or AI visibility software. That does not make those figures wrong, but it does mean Pew is the one you can put in front of a sceptical board.

What Pew actually measured

Real browsing behaviour, passively metered — not a keyword index and not a survey. 900 US adults who consented to an app recording their browsing, producing 68,879 unique Google searches in March 2025, of which 12,593 produced an AI summary.

The attribution people get wrong

KnowledgePanel Digital is operated by Ipsos, not Pew. Pew’s own probability panel is the American Trends Panel. Write ‘Ipsos’s KnowledgePanel Digital, used by Pew’ if you are being careful.

The limitation worth stating

It is US-only, March 2025, and 900 people. It is the highest-provenance number available and it is still a snapshot of one month in one country.

Why does every source give a different AI Overview trigger rate?

Because each one measures a different denominator. Published figures for roughly the same period run from about 16% to about 48%, and none of those numbers is wrong — they are answers to different questions.

Ahrefs’ 20.5% is 20.5% of the SERPs in Ahrefs’ own keyword index, which is long-tail heavy and not weighted by search volume. Conductor’s 25.11% is a share of its enterprise customers’ keyword set. BrightEdge tracks commercial verticals, which skew high by design. Semrush tracks a different panel again. Pew’s 18% is the only figure derived from what real people actually searched.

The practical consequence is that you should never quote a single trigger rate as though it described Google. Quote the one whose denominator matches your question, and say what the denominator was.

AI Overview trigger rate by measurement — These are not contradictory. Each denominator is a different sample: an SEO tool’s index, an enterprise keyword set, a commercial-vertical tracker, or real metered browsing.

Why the trigger-rate figures differ — 1 = no, 3 = yes. Only Pew measures what real people actually searched; the others measure keyword indexes.

Keyword index — Denominator. What an SEO tool tracks.

Enterprise keyword set — Denominator. One vendor's customers.

Commercial verticals — Denominator. Deliberately skewed high.

Metered browsing — Denominator. What people actually searched.

Clickstream panel — Denominator. Panel composition matters.

Self-reported — Denominator. The company's own count.

What each trigger-rate denominator actually is
Source	Denominator	What it can tell you	What it cannot
Pew	Real metered user searches	What people actually encountered	Anything after March 2025, or outside the US
Ahrefs	Its own keyword index, desktop	Relative prevalence across a large SERP sample	Share of real searches, since it is not volume-weighted
Conductor	Enterprise customer keywords	What enterprise brands’ terms look like	Consumer or long-tail behaviour
BrightEdge	Commercial verticals	Where trigger rates are highest	A general Google average
Semrush	A tracked keyword panel	Trend over time on a stable set	Anything about keywords outside the panel

How much does the trigger rate vary by industry?

By roughly eleven times in Conductor’s data — Health Care at 48.75% against Real Estate at 4.48% — which makes any single headline percentage useless for a specific sector.

This is the most practically important statistic on the page and the least quoted. If you are in healthcare, the general figure understates your exposure by half. If you are in real estate, it overstates it fivefold.

The direction is firmer than any individual number. seoClarity’s 500-million-keyword dataset reproduces the same ordering with a wider spread — healthcare around 88%, real estate under 3% — and BrightEdge also ranks healthcare highest. Every independent dataset found agrees on the ranking even where the magnitudes differ.

AI Overview trigger rate by industry — seoClarity’s 500M-keyword dataset reproduces the same direction with a wider spread (Healthcare ~88% vs Real Estate under 3%). Every dataset found ranks healthcare highest and real estate lowest.

AI Overview trigger rate by industry, Conductor
Industry	Trigger rate
Health Care	48.75%
Financials	25.79%
Utilities	25.40%
Consumer Staples	6.82%
Real Estate	4.48%

Why healthcare is highest

Health questions are informational, high-volume and well-suited to a summarised answer, and they are exactly the queries where a search engine has most incentive to answer directly.

Why real estate is lowest

Property search is transactional and local, and the useful result is a listing rather than a paragraph. There is less for a summary to do.

Why the direction matters more than the number

Three independent datasets with different denominators all rank healthcare highest and real estate lowest. That agreement is stronger evidence than any single percentage from any one of them.

What to do with it

Find your own sector’s figure before planning around a general average, and treat the general average as close to meaningless for planning purposes.

What about the enormous user numbers — 2.5 billion, 1 billion, 800 million?

They are real disclosures, but they are first-party statements with no methodology attached, and two of them are routinely relabelled into something they do not say.

Alphabet’s figures come from an investor presentation and earnings remarks. Alphabet never defines what an active user is for these surfaces, never states the measurement window, never discloses cross-device de-duplication, and none of the figures appear in audited financials. That does not make them false. It means they cannot be checked, and they were published by the party justifying a very large infrastructure spend.

AI Overviews 2.5B — First-party. Alphabet, June 2026, no method.

AI Mode 1B monthly — First-party. Alphabet, May and July 2026.

Gemini 950M MAU — First-party. Alphabet, July 2026; later superseded.

No 'active user' definition — Caveat. Never published by Alphabet.

No measurement window — Caveat. 28-day or 30-day never stated.

Not in audited financials — Caveat. Investor presentation, not accounts.

First-party user figures, and what they actually say
Figure	Exact wording	Source and date	The catch
AI Overviews	‘AI Overviews now have over 2.5 billion users each month’	Alphabet investor presentation, 3 June 2026	‘Users each month’ is not ‘monthly active users’. It most likely counts anyone served a page containing an AI Overview — an impression, not an intentional visit
AI Mode	‘surpassed 1 billion monthly active users’	Google I/O 19 May 2026; restated on the Q2 earnings call 22 July 2026	No definition of active user, no window stated
Gemini app	‘950 million monthly active users, with daily active users tripling’	Alphabet Q2 2026 earnings remarks, 22 July 2026	Superseded within weeks by reporting of 1 billion — any figure cited to Q1 2026 is a misattribution, because Q1 contains no consumer user figure at all
ChatGPT	No absolute figure published in OpenAI’s own adoption dataset	OpenAI adoption release, 30 June 2026	Every series is indexed to a July 2023 baseline. The 800M figure traces to Sam Altman’s DevDay keynote of 6 October 2025, not to any OpenAI dataset

The relabelling that matters most

Alphabet wrote ‘users each month’ about AI Overviews. A great many articles report that as ‘2.5 billion monthly active users’. Those are materially different claims: one counts people who were served a page containing an AI Overview, the other implies deliberate use of a product. AI Overviews appear inside a product that already has billions of users, so the figure measures exposure rather than adoption.

Why AI Mode and Gemini must never be added together

Overlap between them is undisclosed. Summing them produces a number with no meaning.

What OpenAI actually publishes

Change relative to a July 2023 baseline, as a multiple. The dataset contains no absolute weekly or monthly user count anywhere. The absolute figures in circulation come from keynote and press disclosures, which is a different kind of evidence and should be attributed that way.

How to cite these honestly

Write ‘Alphabet reported…’ rather than ‘AI Mode has…’. The first is checkable and true; the second presents an unaudited company claim as an established fact.

How a statistic drifts — This is how ‘users each month’ became ‘monthly active users’ — a materially different claim.

Want this done for your site?We build and maintain the search, content and paid programmes described on this page.

Get a free proposal

Which widely-quoted statistics could not be traced at all?

Three of the most repeated: ChatGPT’s user count as an OpenAI-published figure, Google’s ‘billions of clicks to websites every week’, and the claim that Reddit is the most-cited domain in AI answers.

Untraced does not mean false. It means that after following each one back through the chain of citations, no primary measurement with a stated sample, method and date was found. Publishing them as facts is how the drift continues.

800M weekly ChatGPT users — Untraced to OpenAI. From a keynote, not a dataset.

'Billions of clicks weekly' — Untraced. No primary measurement found.

Reddit most-cited domain — Untraced. No dataset with a stated sample found.

Perplexity user figures — Untraced. No audited metric published.

Citation source composition — Untraced. No verified dataset exists.

Tracker accuracy comparisons — Untraced. Nobody has published one.

Widely-quoted figures that could not be traced to a primary measurement
Claim	Where it actually comes from	Why it does not hold as stated
‘800 million weekly ChatGPT users’ as an OpenAI statistic	Sam Altman, DevDay keynote, 6 October 2025	OpenAI’s own adoption dataset publishes no absolute user count at all
‘AI Overviews drive billions of clicks to websites every week’	Attributed to Google	No primary measurement found; the surrounding earnings language carries no number, sample or window
‘Queries are at an all-time high because of AI Overviews’	Alphabet Q1 2026 earnings remarks	Total query counts rise over time regardless, so the claim is near-unfalsifiable, and it is a causal attribution from the beneficiary with no data attached
‘Reddit is the most-cited domain in AI answers’	Widely repeated, origin unclear	No dataset with a stated crawl, query sample and date was found
Perplexity user or referral numbers	Company statements and panel estimators	Perplexity publishes no audited user metrics
Share of AI citations by source type	No verified dataset	The nearest claim measures organic-rank overlap, which is a different question

What has nobody measured yet?

Perplexity’s real audience, the composition of AI citations by source type, and whether any two AI visibility trackers agree about the same brand on the same day.

The third is the most useful gap, because the entire AI visibility tooling category depends on the assumption that its measurements are stable. Nobody has published a controlled comparison. Until somebody does, every tracker’s numbers are internally consistent and externally unvalidated.

The second gap matters commercially. A great deal of advice is built on assumptions about where AI answers draw their citations from — your own site, third-party editorial, forums, review platforms. No primary dataset with a stated method was found for any of it.

What nobody has actually measured — Bad = no primary dataset with a stated sample and date was found. These are open questions, not settled ones.

How should you read any AI search statistic?

Four questions, in order. Most statistics in this field fail at the second one.

How to read any AI search statistic — Most statistics in this field fail at question two.

Who published it first? — Question. Follow it past the blog.

What is the denominator? — Question. 20% of what, exactly?.

When was it measured? — Question. This field moves quarterly.

Who profits from belief? — Question. Context, not disqualification.

Is the sample stated? — Question. No number, no confidence.

Is it a measurement or a claim? — Question. Earnings calls are claims.

Who published it first? Follow it back past the blog you found it on — usually two or three hops.
What is the denominator? 20% of what? An SEO tool’s keyword index is not the same thing as real searches.
When was it measured, and when was it published? Those are different dates and both matter in a field moving this fast.
What was the sample size, and is it stated at all? No number, no confidence.
Is it a measurement or a claim? An earnings call is a claim. A metered browsing panel is a measurement.
Who profits if you believe it? Not disqualifying — Ahrefs and Conductor both publish real method — but it belongs in how you weight it.
Has it been relabelled? ‘Users each month’ becoming ‘monthly active users’ changes the claim.
Does an independent dataset reproduce the direction? Agreement on direction is stronger than any single figure.

Question one: who published it first

The blog you are reading almost never measured anything. Two or three hops back you will usually find either a vendor report with a method, or a keynote with none.

Question two: what is the denominator

This is where most of these figures fail. A trigger rate measured across an SEO tool’s keyword index answers a different question from one measured across real user searches, and the two differ by a factor you cannot predict.

Question three: when

AI surfaces change quarterly. A September 2025 figure is a real measurement of September 2025 and a guess about today.

Question four: who benefits

Vendors selling AEO software benefit from high trigger rates and high disruption. Google benefits from the opposite. Both publish. Read both, and note which one attached a method.

The relabelling check

Compare the wording in the quote to the wording in the original. Impressions become active users, SERPs become searches, and a keynote becomes a dataset.

The direction test

Where three datasets with different denominators agree on a direction — as they do on healthcare having the highest AI Overview trigger rate — that agreement is worth more than any one of their numbers.

Want to know your own numbers rather than the industry’s?

Industry averages cannot tell you whether AI answers cite you. A fixed-scope audit measures your own crawler access, rendering and citation baseline — and hands you the prompt set so you can re-run it yourself.

/ai-visibility-audit

Why we published this instead of another statistics listicle

Because the existing ones repeat each other. Almost every AI SEO statistics article we checked carries figures without a date, without a denominator, and without a link to whoever measured them first — and several carry numbers the original publisher does not state.

This page is the same exercise done backwards: start from the claim, follow it to the publisher, read the method, and report what is actually there including the gaps. Where a figure could not be traced, it is listed as untraced rather than quietly dropped, because the absence is itself the finding.

Our interest: Progression Agency sells answer engine optimisation services. A page concluding that AI search is overhyped would suit that interest less well than this one does, which is the reason to state the interest rather than hide it.

What we did

Ran 107 verification passes across the source set, each attempting to refute the claim against the original publisher’s page rather than confirm it.

What we did when verification failed

Went back to the primary source and read it directly. The Pew click-through figures are on this page because we re-read them off Pew’s own page after the automated pass could not re-confirm them within budget.

What we did not do

Commission any original research, run our own crawl, or accept a vendor briefing. Nothing on this page rests on a measurement we made.

What would improve this page

A controlled comparison of visibility trackers, a primary dataset on citation source composition, and any audited Perplexity figure. If you have one, we will cite it.

A controlled comparison of visibility trackers, a primary dataset on citation source composition, and any audited Perplexity figure.

Is AI search actually replacing Google?

Not on any traced measurement. Conductor put AI referral traffic at 1.08% of all website traffic across 1,215 enterprise domains — real, growing, and small. The replacement narrative runs well ahead of the data in both directions.

What the data supports

That AI Overviews are prevalent on a substantial minority of searches, that their prevalence varies enormously by sector, and that their presence coincides with materially fewer clicks on traditional results.

What the data does not support

That AI assistants have replaced search traffic. One per cent of sessions is not a replacement, and nobody has published a measurement showing otherwise.

What Google says, and why it is not evidence

That queries are at an all-time high and AI Overviews drive Search growth. No number, sample or window is attached, total query counts rise over time regardless, and it is a causal claim from the party that profits from it.

Why both sides overstate

Vendors selling AEO benefit from disruption. Google benefits from continuity. Neither has published the study that would settle it.

What should you actually do with these numbers?

Use them to size the question, not to answer it. Industry averages cannot tell you whether AI answers cite your business — only measuring your own prompts can.

Find your sector’s figure, not the average

The eleven-fold industry spread means the general trigger rate is close to useless for planning.

Measure your own baseline

A fixed prompt set re-run monthly tells you more about your business than every statistic on this page combined.

Do not plan around a single percentage

Especially not one whose denominator you cannot name.

Re-check the figures quarterly

Every measurement here will be historical within a year.

A fixed prompt set re-run monthly tells you more about your business than every statistic on this page combined.

How do vendor studies and academic studies differ here?

Vendor studies have far larger samples and a commercial stake. The one non-commercial study has a small sample and no stake. Both are usable; neither is sufficient alone.

Sample size favours the vendors

Ahrefs measured 146 million SERPs. Pew measured 68,879 searches. On raw scale there is no contest.

Provenance favours Pew

But Ahrefs measured its own index, and Pew measured what people actually did. Scale and representativeness are different virtues.

Use them for different jobs

Vendor data for relative patterns and trends; Pew for the headline claim you need to defend.

Neither is a substitute for your own numbers

Both describe an average you do not belong to.

Not sure which of these applies to you?Tell us the situation and we will say plainly what we would do first, and what we would not.

Talk it through

What changed between 2025 and 2026?

The user figures grew, the trigger-rate measurements stayed in the same broad range, and the quality of the statistics circulating got worse rather than better as repetition outpaced measurement.

User numbers grew and got vaguer

Gemini went 400M to 750M to 900M to 950M across five quarters of disclosures, none with a published definition of an active user.

Trigger rates stayed broadly stable

The 16-48% spread reflects sampling differences more than change over time.

The drift got worse

‘Users each month’ becoming ‘monthly active users’ is now near-universal in coverage.

The gaps did not close

Perplexity figures, citation composition and tracker accuracy were unmeasured in 2025 and remain unmeasured.

Perplexity figures, citation composition and tracker accuracy were unmeasured in 2025 and remain unmeasured.

Who measured the click-through impact, and can it be trusted?

Pew Research Center, in March 2025. It is the only click-through study in this field produced by an organisation that does not sell SEO or AI visibility software.

What is the single most useful number on this page?

The industry spread: Health Care 48.75% against Real Estate 4.48%. Everything else is an average you probably do not belong to.

Which figures are safe to put in a board deck?

Pew’s click-through numbers and Conductor’s referral share, both with the date and sample stated. Avoid anything from an earnings call unless you attribute it as a company claim.

Are the AI Overview trigger rates rising over time?

The published measurements do not establish that cleanly, because the spread between vendors is wider than the movement between periods. Comparing Ahrefs in September 2025 to BrightEdge in early 2026 compares two samples, not two dates.

How many AI search statistics did you check?

Every figure in the source set, across 107 verification passes, each attempting to refute the claim against the original publisher rather than confirm it. Five survived with a stated method, three are unverifiable disclosures, and three could not be traced at all.

What would change our conclusions?

A controlled comparison of AI visibility trackers, an audited Perplexity figure, or a primary dataset on citation source composition. Any of the three would move this page materially, and we will cite them if they appear.

Can we use these statistics in our own marketing?

Yes, with the attribution and date carried across. The reason this page exists is that most reuse drops both — and a statistic without a denominator is not evidence, it is decoration.

The reason this page exists is that most reuse drops both — and a statistic without a denominator is not evidence, it is decoration.

Why does the wording of a statistic matter so much?

Because relabelling changes the claim. ‘Users each month’ counts people served something; ‘monthly active users’ implies deliberate use. One word of drift turns an impression count into an adoption figure.

What this page is built on — Research run 15-16 September 2026.

Frequently asked questions

What percentage of Google searches show an AI Overview?

There is no single answer. Published measurements for roughly the same period run from about 16% to 48%, because each vendor samples a different keyword universe. Pew’s 18% is the only figure derived from real metered user searches; Ahrefs measured 20.5% of 146,122,391 desktop SERPs in its own index.

Do AI Overviews reduce clicks to websites?

Pew Research Center measured users clicking a traditional search result in 8% of visits where an AI summary appeared, against 15% where one did not — nearly twice as often without. That is 68,879 real searches from 900 metered US adults in March 2025.

How often do people click a link inside an AI summary?

In 1% of visits, according to Pew’s March 2025 browsing data.

Which AI search statistic is the most trustworthy?

Pew Research Center’s click-through study. It is the only measurement in this literature from an institution with no commercial interest in the result, and it uses passively metered browsing rather than a keyword index.

How many people use ChatGPT?

OpenAI’s own published adoption dataset contains no absolute user figure — every series is indexed to a July 2023 baseline. The widely-quoted 800 million weekly figure traces to Sam Altman’s DevDay keynote of 6 October 2025, which is a company statement rather than a published dataset.

How many people use Google’s AI Overviews?

Alphabet stated ‘over 2.5 billion users each month’ in its June 2026 investor presentation. Note the wording: ‘users each month’ is not ‘monthly active users’, and it most likely counts anyone served a page containing an AI Overview rather than deliberate use.

How many people use AI Mode?

Alphabet reported surpassing 1 billion monthly active users, first at Google I/O in May 2026 and again on the July 2026 earnings call. No definition of active user or measurement window is published.

Can I add AI Mode and Gemini user numbers together?

No. Overlap between them is undisclosed, so the sum has no meaning.

Which industry sees the most AI Overviews?

Health Care, at 48.75% in Conductor’s data, against Real Estate at 4.48% — a roughly elevenfold spread. seoClarity’s much larger dataset reproduces the same ordering with an even wider gap.

How much traffic comes from AI referrals?

Conductor measured 1.08% of all website traffic — 35.7 million of 3.3 billion sessions across 1,215 enterprise customer domains, May to September 2025. That measures enterprise brand sites rather than publishers.

Is Reddit really the most-cited domain in AI answers?

We could not trace that claim to any dataset with a stated crawl, query sample and date. Treat it as untraced.

How many people use Perplexity?

Perplexity publishes no audited user metrics. Figures in circulation come from company statements and third-party panel estimators, neither of which we could trace to a primary measurement.

What share of AI citations come from a brand’s own site?

No verified dataset exists. The nearest available claim measures overlap with organic top-ten rankings, which is a different question.

Why do different sources disagree so much?

Because they measure different denominators — an SEO tool’s keyword index, an enterprise keyword set, a commercial-vertical tracker, or real metered browsing. They are answers to different questions, not contradictory answers to one.

Should I trust statistics from SEO software companies?

They are usable when the method is stated, and Ahrefs and Conductor both state theirs. Weight them knowing the publisher benefits if you believe AI search is disruptive.

Should I trust Google’s own figures?

As disclosures, yes. As measurements, they cannot be checked — no definition of an active user, no window, no de-duplication, and not in audited financials.

How current are these statistics?

The measurements range from March 2025 to mid-2026 and each is dated on this page. This field moves quarterly, so treat any figure older than a year as historical.

What does ‘users each month’ mean versus ‘monthly active users’?

They are materially different. The first can count anyone served a page; the second implies deliberate use. Alphabet used the first phrasing for AI Overviews and much of the coverage reports the second.

Has anyone compared AI visibility trackers for accuracy?

Not that we could find. It is the most useful missing study in this field, because the whole tooling category rests on the assumption that its measurements are stable.

How do I check a statistic myself?

Follow it back to the original publisher, find the denominator, check the date, and see whether a sample size is stated. Most fail at the denominator.

What is the difference between a measurement and a claim?

A measurement states a sample, a method and a window. A claim does not. Earnings calls produce claims; metered browsing panels produce measurements.

Do you have a commercial interest in these numbers?

Yes, and it is worth stating. Progression Agency sells answer engine optimisation services, so we benefit if you believe AI search matters. That is why every figure here is attributed and dated so you can check it.