The Source-to-Mention Gap: 205 GTM Buyer Queries Inside Google's AI Overviews

Cleanlist published 205 B2B data and GTM buyer queries pulled on Google US desktop on September 1, 2026. Six returned no SERP at all, leaving 199 usable. An AI Overview appeared on 196 of those 199, which is 98.5%. Only 73 of those 196 answers returned a reference list, and that number governs most of this page. Across the answers that did expose sources, the most-cited domain is youtube.com on 37, then pipeline.zoominfo.com on 27, reddit.com on 23, apollo.io on 16 and cleanlist.ai on 14. The brands the answers actually name are Apollo on 47, ZoomInfo on 40, Clay on 32, Cognism on 25, Lusha on 18 and Cleanlist on 2. Google reads us fifth and names us fifteenth.

Last updated: September 1, 2026. Every SERP in this study was collected on September 1, 2026.

How often does Google show an AI Overview on B2B data and GTM buyer queries?

On 196 of the 199 usable SERPs, which is 98.5%. Against all 205 published queries the figure is 95.6%, and the difference is worth stating plainly: 6 pulls returned nothing at all, and calling those "no AI Overview" would understate coverage. The three usable SERPs that genuinely lacked an AI Overview on September 1, 2026 were "hubspot data enrichment", "uplead pricing" and "best mobile number data provider". For a GTM leader the practical reading is that the blue-link SERP has effectively stopped existing in this category. On the queries your buyers type before they ever see your pricing page, an answer is written above the results roughly nineteen times in twenty, and the three exceptions here look like noise rather than a pattern.

Which domains do Google's AI Overviews cite most on GTM queries?

Counting each domain once per AI Overview, and dividing by the 73 answers that returned a reference list, the leaderboard is youtube.com 37 (50.7%), pipeline.zoominfo.com 27 (37.0%), reddit.com 23 (31.5%), apollo.io 16 (21.9%), cleanlist.ai 14 (19.2%), cognism.com 12 (16.4%), clay.com 11, salesmotion.io 11, syncgtm.com 11, autobound.ai 10 and salesforce.com 9. Expressed against all 196 AI Overviews, including the 123 stubs that could never cite anybody, the same figures read youtube.com 18.9%, pipeline.zoominfo.com 13.8%, reddit.com 11.7% and cleanlist.ai 7.1%. Rolling subdomains up to the registrable domain moves ZoomInfo to second with 29 and Apollo to fourth with 18, and leaves cleanlist.ai fifth on either basis. Cleanlist is the only company in the top five that is not a household name in this category, which is the whole reason this study exists.

Which vendor brands do AI Overviews actually name?

Naming is a different leaderboard, and it is the one that sends a buyer to a signup page. Across the 73 answers that carried a readable body: Apollo 47 (64.4%), ZoomInfo 40 (54.8%), Clay 32 (43.8%), Cognism 25 (34.2%), Lusha 18 (24.7%), Instantly 14 (19.2%), Clearbit 5 (6.8%), then FullEnrich, Hunter, Seamless, Snov and UpLead on 4 each (5.5%), Findymail and RocketReach on 3 (4.1%), Bombora, Cleanlist, Prospeo and Smartlead on 2 (2.7%), and Kaspr and Skrapp on 1 (1.4%). Twenty vendor names appear inside this corpus. Cleanlist sits in a four-way tie for 15th. Answers that name anyone name a median of 3 vendors, with a maximum of 7, so the shortlist Google writes is short.

Why does every rate on this page divide by 73 instead of 196?

Because a mention can only come from an answer that has text, and 123 of the 196 AI Overviews in this corpus have none. This is the correction that changed the most numbers on this page, so the evidence for it is worth showing rather than asserting.

Of the 73 AI Overviews that returned a reference list, 67 name at least one vendor, which is 92%. If the 123 stubs contained readable answer text that our detector simply found no brands in, we would expect roughly 113 of them to name somebody. The observed number is zero. Not one of the 123 carries a single brand mention.

That pattern holds well beyond this corpus. Across the fifteen raw SERP files retained for this sprint, which include re-pulls of overlapping query sets, there are 425 AI Overviews with no reference list, and zero of them carry a brand mention. A detector that finds brands in 92% of answers with a body and in 0% of answers without one is not failing to read those answers. There is nothing there to read. The collector requested the AI Overview element without setting the API's asynchronous flag, and Google loads AI Overview content after the initial SERP response, so the element was detected and its content was not.

Dividing by 196 would therefore mix two populations: answers that could have named a vendor and answers that could not. It makes every vendor look less visible than it is, ours included, and it makes the study's central comparison unreliable. The counts themselves are unaffected, because a stub contributes zero to both sides.

What is the source-to-mention ratio, and why does it matter?

We are naming this metric because the two leaderboards above disagree and nobody was measuring the disagreement. The source-to-mention ratio is a vendor's brand mentions in AI Overview text divided by the number of AI Overviews that cite its own domain. Above 1.0, Google names you more often than it reads you, which means third parties are carrying your name into answers you never appear in. Below 1.0, Google reads you and writes about somebody else. Apollo scores 2.61 (47 mentions against 18 citations). Cleanlist scores 0.14 (2 mentions against 14 citations). That is an 18.3x gap, and on the stricter exact-hostname basis, where Apollo counts 16 citations, it widens to 20.6x. The ratio separates the vendors Google reads from the vendors Google recommends.

Which vendors convert citations into mentions, and which do not?

The full table, ranked by ratio, with citations rolled up to the registrable domain:

Ratios built on one, two or three citations are unstable, so read the top three rows as directionally interesting rather than precise. Restricted to the six vendors with at least five citations, the order is Clay 2.91, Apollo 2.61, Cognism 2.08, ZoomInfo 1.38, UpLead 0.57 and Cleanlist 0.14. We finish last of six by a factor of four. Clay's row carries one contaminated mention described in the corrections section; on the adjusted count it reads 31 mentions and a ratio of 2.82, which does not change its position.

Does being cited by an AI Overview get your brand named in it?

No, and this is the most useful number in the study. There are 217 brand-mention events across the corpus. In 72 of them (33.2%) the named vendor's own domain is also cited in that same answer. In 145 of them (66.8%) it is not. Five vendors are named without their domain ever being cited anywhere in the corpus: Clearbit 5 times, Bombora 2, Prospeo 2, Smartlead 2 and Kaspr 1. Apollo is named in 47 answers and cited in only 17 of them, so 30 of its mentions arrive with no apollo.io page in the reference list. Getting your own page into the source list is neither necessary nor sufficient for being named. Something on other people's pages is doing the work.

Why is Cleanlist cited 14 times and named twice?

The 14 answers citing cleanlist.ai are "b2b data enrichment tools", "b2b contact database", "data enrichment platform", "email enrichment tool", "salesforce data enrichment", "contact enrichment software", "gtm engineering", "clay alternatives", "rocketreach pricing", "clearbit pricing", "apollo vs zoominfo", "apollo vs lusha", "seamless ai pricing" and "data enrichment tools". Only two of them, "email enrichment tool" and "clay alternatives", name Cleanlist. In the other 12, Google quotes our benchmark work and then recommends Apollo, ZoomInfo, Clay, Cognism and Lusha in the sentence next to it. We publish original enrichment testing, most recently the 500-Lead Enrichment Benchmark, and that research is doing its job as a source. It is not doing anything for the brand, because a source and a subject are different roles in an answer.

What distinguishes the vendors Google names from the ones it only reads?

We can test three candidate explanations against this corpus, and one of them survives better than the others. Correlating each of the 20 vendors' mention counts, Spearman rank correlation gives 0.86 for the number of distinct third-party domains ranking in these SERPs with a page title that names the vendor, 0.67 for the vendor's own AI Overview citations, and 0.51 for the vendor's own organic top-10 presence. The ordering is consistent with corroboration across independent sources mattering more than anything on the vendor's own property. This is correlation across 20 vendors on one snapshot, not causation, and the strongest measure has a serious circularity problem described in the next passage.

How circular is the third-party naming measure?

Very, and pretending otherwise would break the study. Thirty-two of the 199 usable queries name a vendor in the query string itself ("clay alternatives", "apollo pricing", "apollo vs zoominfo" and so on), and those queries obviously surface pages titled "14 Best Clay Alternatives". If you exclude every query whose own text names the vendor, the distinct-third-party-title counts collapse to 4 for Apollo, 3 for Hunter, 3 for Instantly, 2 for ZoomInfo, 2 for Clay, 1 for Lusha and 0 for everyone else, and the Spearman correlation falls to 0.72 on numbers too small to lean on. Two of Instantly's three surviving domains are adverb false positives rather than vendor references, which is the same detector limit flagged for that row throughout. Cleanlist scores 0 either way: across 199 SERPs and 1,987 organic results, not one third-party page title in this corpus contains the word "Cleanlist". That zero is a real finding even though the correlation built on it is weak.

Does ranking in the organic top 10 get you named in the AI Overview?

Not on this evidence. cleanlist.ai ranks in the organic top 10 on 30 of the 199 usable SERPs, which is more than any vendor except ZoomInfo (86), Cognism (45) and Apollo (36), and it is named twice. Lusha ranks on 6 and is named 18 times. Clearbit ranks on 2, is cited zero times, and is named 5 times. Kaspr ranks on 17 and is named once. Snov ranks on 18 and is named 4 times. Our own median organic position across those 30 queries is 7, with two number-one rankings. Classical ranking and answer-text naming are close to decoupled in this category, which is uncomfortable if your entire GTM motion is an SEO motion. Ours largely was.

Is YouTube really the most-cited domain in GTM AI Overviews?

Yes, on this corpus, and it is not close. youtube.com is cited in 37 of the 73 AI Overviews that exposed a reference list, which is 50.7%, and it sits 10 citations clear of the second-place domain. It is also cited well beyond its organic footprint: youtube.com ranks in the organic top 10 on 39 of the 199 usable SERPs, but of the 37 AI Overviews citing YouTube, only 12 had a YouTube result in that query's own top 10. The honest limit on this finding is that it shows citation, not causation, and certainly not conversion. We did not measure whether a YouTube citation sends anyone anywhere. What it does establish is that a GTM category with almost no video program is leaving the single largest citation surface uncontested.

How often does Reddit rank in the organic top 10 for these queries?

reddit.com appears in the organic top 10 on 143 of the 199 usable SERPs, which is 71.9%. Counting business.reddit.com as well brings it to 144, or 72.4%. Where Reddit ranks, its best position has a median of 3, and it sits in the top three on 97 of those 144 SERPs, or 67.4%. As an AI Overview source it is cited on 23 of the 73 answers with a reference list (31.5%), third behind YouTube and ZoomInfo's pipeline subdomain. Forum and social domains together (Reddit, Quora, LinkedIn, Medium, Hacker News) account for 25 of the 505 citations, or 5.0%, so Reddit's dominance is a ranking phenomenon first and a citation phenomenon second. Your buyer sees a Reddit thread above the fold on roughly seven queries in ten.

How concentrated is AI Overview citation, or is there a long tail?

Both, and the shape is the point. There are 505 domain-and-answer citation pairs across 208 distinct hostnames. The top 10 hostnames account for 172 of the 505, which is 34.1%. The top 5 hold 23.2%, the top 20 hold 45.9% and the top 50 hold 62.6%. At the other end, 127 of the 208 hostnames (61.1%) are cited exactly once in the entire corpus. Answers that exposed a reference list cited a median of 7 domains, mean 6.9, maximum 12. So there is a stable handful of domains Google keeps returning to, and then a very wide field of single-appearance sources. A new domain can enter that field on one good page, and almost none of them ever leave it for the head.

Do AI Overviews cite pages that already rank organically?

Mostly yes, which is the good news for anyone with an SEO program. Across the 73 AI Overviews that exposed a reference list, an average of 54.6% of the domains cited in the answer also appear in that query's own organic top 10. Seventy-one of the 73 cite at least one domain that also ranks. The AI Overview is drawing heavily from the same page-one pool rather than from some separate index, so ranking remains the entry ticket. It is just not the whole game: as the passage above shows, the domains that get read and the brands that get named come apart badly once you are inside the answer. Rank to get cited. Get written about elsewhere to get named.

Which query types trigger an AI Overview, and which expose a source list?

Using the published regex, and the 199 usable SERPs as the base: alternatives 10 of 10 (100%), comparison_vs 5 of 5 (100%), how_to 17 of 17 (100%), other 80 of 81 (98.8%), best_x 62 of 63 (98.4%) and pricing 22 of 23 (95.7%). The spread is small enough that the practical answer is "every type". The type mix of the 205 published queries was other 84, best_x 64, pricing 24, how_to 18, alternatives 10 and comparison_vs 5, so the small-n classes carry wide confidence intervals and a 100% on five queries should not be read as a law. If you were hoping some corner of the buyer journey still gets you a clean ten blue links, this corpus does not find it.

Whether the answer shows its sources varies enormously by type, and it is a better planning signal than the trigger rate. Of the AI Overviews in each class, the share that returned a reference list was: alternatives 9 of 10 (90.0%), comparison_vs 3 of 5 (60.0%), other 37 of 80 (46.3%), pricing 8 of 22 (36.4%), best_x 15 of 62 (24.2%) and how_to 1 of 17 (5.9%). Comparison and alternatives answers show their work. Instructional answers almost never do. For a GTM marketer that reorders the content plan: the pages with the best odds of appearing as a visible, clickable source are the comparison and alternatives pages, and a how-to has to earn its return some other way, because on this evidence Google rarely credits anyone for it. The exposure rates also carry the widest error bars on this page, because comparison_vs rests on 5 answers and alternatives on 10.

What did this study get wrong, and what changed before publication?

Five things, and all five moved numbers that had already been circulated internally. They are listed here with the old figure, the new figure and the reason, because a study that publishes only its final state is asking to be trusted rather than checked.

1. The mention denominator was wrong, and it was the biggest error on the page. An earlier draft divided every mention count by 196, the number of AI Overviews, and published Apollo at 24.6%, ZoomInfo 20.6%, Clay 16.1%, Cognism 13.1%, Lusha 10.1%, Instantly 7.0% and Cleanlist 1.0%. Those rates are wrong. Only 73 of the 196 answers returned a body that a brand name could appear in, so the correct rates divide by 73: Apollo 64.4%, ZoomInfo 54.8%, Clay 43.8%, Cognism 34.2%, Lusha 24.7%, Instantly 19.2% and Cleanlist 2.7%. The evidence that the other 123 are empty rather than silent has its own passage above, and the short version is that zero of them carry a brand mention while 92% of the answers with a body do. What this changes: every mention rate, roughly by a factor of 2.7. What it does not change: any count, any ratio, any ranking, and the finding that Cleanlist is read fifth and named fifteenth. Both sides of the ratio are counts, so the ratio was never affected.

2. The citation counting unit was wrong in an earlier pass. That pass counted every appearance of a domain inside one AI Overview's reference list as a separate citation. On this corpus that method returns 563 citation instances instead of 505, inflating YouTube from 37 to 74, Reddit from 23 to 27, apollo.io from 16 to 18, autobound.ai from 10 to 12 and salesforce.com from 9 to 11. Expressed as a rate over the 73 answers that carried a reference list, the buggy method gives YouTube 74 divided by 73, which is 101.4%. A share above 100% is the signature of counting the wrong unit. An earlier pass on a smaller query set published 93.9%; we could not reproduce that exact figure here and are not going to pretend we could. The corrected figure is 50.7% of the 73 answers that exposed sources.

3. One row is a homonym, and it is not the row we originally flagged. The AI Overview for "clay pricing" is about pottery. Its references are lagunaclay.com, artclayworld.com, theceramicshop.com, facebook.com, google.com and reddit.com, and our detector scored the word "Clay" in it as a mention of the GTM vendor. The row's organic top 10 is entirely Clay.com content, which is why it looked clean from the ranking side. It is the same failure mode we had already flagged for Instantly, on a larger row. We are keeping the row and publishing the adjusted figures alongside every count it touches: Clay 32 mentions adjusted to 31 and its ratio 2.91 adjusted to 2.82, reddit.com 23 citations adjusted to 22, 505 citation pairs adjusted to 499, 208 hostnames adjusted to 205, and 127 hostnames cited exactly once adjusted to 125. Clay stays third on the naming leaderboard either way. Dropping the row entirely would also drop the answer count from 73 to 72 and move every rate on this page by less than a point, which is why we did not do it silently.

4. One published column did not match its own published rule. The dataset's distinct_third_party_domains_with_top10_title_naming_brand field read 1 for Instantly in an earlier version. The rule printed in the reproduction recipe, word-boundary matching on lowercased titles, returns 3: a genuine vendor comparison on devcommx.com titled "Clay vs Apollo vs Instantly: Full Comparison (2026)", plus two adverb false positives on unrelated vendors' own pages, one an email verifier titled "Validate Emails Instantly" and one a CRM enrichment page titled "Fill in CRM gaps instantly". A hand filter had been applied to that one row and was not documented. The dataset now publishes 3, the value the published rule actually returns, with the false positives described in the row's notes. Spearman is 0.86 either way, so no downstream figure moves, but a reproduction recipe that does not reproduce a column is a defect on its own.

5. Three queries were removed for a commercial relationship. The collection pulled 208 queries. Two of them asked for alternatives to a Cleanlist supplier, and a third returned an answer whose only named brand was a supplier. All three were removed before publication, because we do not want to publish a shortlist that ranks a company we buy data from. Every figure on this page and in the dataset is computed on the remaining 205. For the record, this cost the study three AI Overviews that all carried reference lists, so the reference-exposing base fell from 76 to 73 and the citation pairs from 528 to 505. The effect on the headline rates is under a point in both directions: Apollo would read 64.5% instead of 64.4% on the full pull. We are stating that so nobody has to wonder whether the exclusion was chosen to flatter a number.

The study also disagrees with Cleanlist's internal notes in three further places, and our recomputed figures are the ones on this page. Internal notes put AI Overview coverage at 95.6% of the published queries; that is right for the attempted set, and 98.5% of the 199 that actually returned a SERP is the better statistic. Internal notes put Reddit in the organic top 10 on 144 SERPs; the exact-hostname count is 143 of 199 (71.9%), and the 144th is business.reddit.com. Internal notes put apollo.io at 16 citations; 16 is right for the exact hostname, and 18 is right once knowledge.apollo.io and docs.apollo.io are rolled in, which moves Apollo's ratio from 2.94 to 2.61 and the gap against us from 20.6x to 18.3x. Both bases are published above so you can pick.

What are the limitations of this study?

Eight, stated bluntly. One snapshot, one day, one country, one device, one search engine. AI Overviews are personalised and non-deterministic, and this corpus contains no repeat pulls, so we cannot report run-to-run volatility and you should not assume these answers reproduce. Only 73 of the 196 AI Overviews returned a reference list, so every citation figure and every mention figure on this page describes that 37.2% subset and is silent about the other 123. Those 123 are silent because the collector did not set the API's asynchronous AI Overview flag and Google renders the answer body after the initial response, which is a collection defect on our side and the single thing we would fix first on a re-run. Mentions come from a machine parse with a fixed 20-name vocabulary, so any vendor outside that list scores zero by construction; a 21st name, LeadIQ, appears elsewhere in the wider collection and not here. The detector cannot separate a vendor name from an ordinary English word, which costs it in two known places: at least four of Instantly's 14 mentions sit on queries ("data cleansing tools", "data cleaning tools", "what is gtm engineering", "enrichment api") where the adverb is the likelier reading, and one of Clay's 32 sits on a ceramics answer, so treat both rows as upper bounds. The exposure-by-query-type table rests on as few as 5 answers in its smallest class. The query set is ours, not a random sample, and three queries were removed from it for the reason given above. And we are a vendor in our own study.

Where can I download the dataset, and how would somebody reproduce it?

The full open dataset is at /data/gtm-answer-engine-study-2026-09.csv, published under CC BY 4.0. Attribute it to Cleanlist and link back to this page. It is one file with two sections. Section one is 205 rows, one per query, sorted alphabetically. After a blank line, a marker row reading ## SECTION 2 introduces the 20-row per-vendor table with its own header. If your parser dislikes the two-section layout, split the file at the blank line. Every figure on this page is recomputable from these two tables, including the ones where we lose.

Section one column dictionary. query, the search string. query_type, one of comparison_vs, alternatives, pricing, how_to, best_x, other, by the published regex. serp_returned, 1 if the pull returned any results (0 for the 6 failures). has_ai_overview, 1 if an AI Overview element was present. ai_overview_exposed_references, 1 if that AI Overview returned a reference list, which is the flag that selects the 73-row base for every rate. cited_domains, pipe-separated, de-duplicated, lowercased, www. stripped. cited_domain_count. brands_named, pipe-separated vendor names found in the answer text. brand_count. ai_overview_names_cleanlist, 1 or 0. cleanlist_organic_position, 1 to 10, blank if absent. top_organic_domain, the domain at organic position 1. reddit_in_top10, 1 or 0, counting any reddit.com subdomain. youtube_cited_in_ai_overview, 1 or 0. serp_features, pipe-separated list of the SERP feature blocks present.

Section two column dictionary. brand. primary_domain. aio_citations_domain_rollup, AI Overviews citing the vendor's registrable domain or any subdomain of it. aio_citations_exact_hostname, the stricter count. aio_mentions. aio_mention_rate_of_73, mentions as a percentage of the 73 answers that exposed a reference list. source_to_mention_ratio, mentions divided by rollup citations. mentions_with_own_domain_cited and mentions_without_own_domain_cited, which sum to mentions. citations_without_mention. distinct_other_domains_cited_alongside_mentions. distinct_third_party_domains_with_top10_title_naming_brand, by the published word-boundary rule with no hand filtering. organic_top10_queries. queries_whose_text_names_brand, the circularity control. notes, including the supplier relationships and the two homonym caveats.

To reproduce it without us, take column one of the dataset as the query list. Call any live Google SERP API with location_code 2840, language English, device desktop, request the AI Overview element, and set that API's asynchronous AI Overview option, which we did not. For each response that returns a reference list, take the set of distinct reference hostnames, lowercase them, strip a leading www., and count each hostname once per query. That reproduces the domain leaderboard. For the mention counts you need the AI Overview answer text and a vendor vocabulary; ours is the 20 names in section two, and word-boundary matching on lowercase text reproduces every column in section two, including the Instantly value of 3 that an earlier version hand-filtered to 1. Expect your absolute counts to differ from ours, because AI Overviews are regenerated per request, and expect a far higher share of answers to expose references than our 37.2%, because that flag is the one thing we got wrong. What we would expect to survive re-collection is the shape: near-total AI Overview coverage, a fat head of about ten domains, and a source list that disagrees with the brand list.

What should a GTM team actually do with this?

Three things this corpus supports. First, separate the two goals. Ranking a page gets you cited (54.6% of an answer's cited domains also rank on that query), and being written about on other people's pages is what correlates with being named. If your quarterly plan only contains the first, you are building Cleanlist's problem on purpose. Second, look at where the citation surface is thin. YouTube is the most-cited domain in this category and almost nobody in B2B data has a video program. Third, weight comparison and alternatives content, because those answers expose a visible source list 90.0% and 60.0% of the time against 24.2% for "best X" and 5.9% for how-tos, with the caveat that those two classes rest on 10 and 5 answers. That is where a citation is actually visible to a human.

What is Cleanlist doing about its own result?

Publishing this, first. The reason cleanlist.ai gets read 14 times is that we run primary tests and publish the numbers, including the 500-Lead Enrichment Benchmark, where 500 stratified B2B leads returned 98% verified email and 85% direct dial across 25 or more providers, against 70 to 80% email and 30 to 60% phone from a single source. Research is why Google reads us. The 0.14 ratio says research alone will not get us named, and another page on our own domain will not change it. What this corpus points at is third-party coverage: reviews, community threads, video, and other people's comparison posts. We will re-run this exact 205-query corpus with the asynchronous flag set and publish whether that moves the number, including if it does not.