What AI Overviews actually quote, across 7,601 searches

Starting point — I wanted to see how the answer picks

I ran this capture to see how the block works.

The block reads the results, rewrites them into 1 answer, and names the companies it thinks are worth naming. My job is to get a site in front of buyers. So I need 2 things answered. How do you become the company it names? And how much of that is the ranking work I already do?

Well, ask around the industry and you get 2 confident stories that contradict each other. Either the AI block just rewrites the top 3, or it goes off somewhere unknowable and rankings no longer matter. Both camps sound very sure. Neither of them had counted anything.

And a screenshot of 1 search settles nothing either. The same query can show an answer at breakfast and none by lunch. You can’t plan a year on it.

So I collected the data myself and measured it. 3 questions, in plain words: how much of the answer comes out of the results underneath it, which sites it picks, and what makes the difference when 2 pages rank next to each other and the answer takes only 1 of them.

Method — 7,601 searches, matched to their own rankings

7,601 searches inside a single commercial market, taken in 1 capture, 20 results deep. All of them came back complete.

For each search I kept the answer, every source it links and the order it links them in, the top 20 results with their addresses, and every other block on the page. Then I gave each site 2 numbers: a third-party authority score from 0 to 100, and how many of the 7,601 searches it shows up on.

3 cleanup rules, because they move the numbers

First, junk links. 687 of the links inside the answers, 1.7% of them, were not sources at all: 286 pointed back at my own collection endpoint, 365 were search-engine redirect wrappers and shopping cards, and 36 were image thumbnails. All of that mess was sitting in the bucket labeled “came from outside the results” and inflating it by a ninth.

Second, subdomains, dropped on both sides of the comparison. A help center, an app-store listing or a media section sitting on a subdomain is not the brand competing here, and counting it would credit a company for a page it does not really own. That drops 3.0% of links and 3.9% of ranking rows.

Third, the rule that decides half the study. A link can be matched to the ranking 2 ways, and they give different answers. Page level: does the exact address the answer links also rank in the top 20? Site level: does that site rank at all, with any page? Every number below says which one it is.

1 answer, taken apart

So here is a typical answer from the study, names swapped out. 6 links, which is the median, and every counting rule shows up in it at once.

6 links, 5 sites, because Site A appears twice with 2 different pages. 4 of the 6 links point at a page that ranks by its own address. 1 points at a different page of a site that ranks well anyway. 1 comes from a site with no ranking on this search at all.

Every big number on this page is that same breakdown, done 6,985 times and added up.

Finding — it is on almost everything, and it is short

91.9% of the searches had an AI answer. 6,985 out of 7,601. There is nothing left to argue about here. It shows up on 9 searches out of 10.

Now, the median answer runs 182 words and links 6 sources. 118 words at the short end, 250 at the long end. A page competes for 1 of about 6 slots, and those slots sit inside something the length of 2 paragraphs.

What that changes on a plan: a page that used to be 1 of 10 results is now trying to be 1 of 6 sources inside 182 words, and that block sits above everything you used to optimize for.

More links does mean a longer answer, and only just

So yes, more links does come with a longer answer. Triple the links from 3 to 9 and the median goes from 170 words to 199. That is 29 words for 6 extra sources, which is not much of a rule.

Look at the spread instead. Take the 1,205 answers that link exactly 5 sources. A quarter of them run under 143 words, a quarter run over 212, and the ends are 123 and 240. Same number of sources, double the writing. The link count is not what sets the length.

My read is that Google decides how much explaining a question needs, and that call comes off the question itself. The data half agrees. Length lines up with the query’s own length and with the number of other blocks on the page about as well as it lines up with the link count. All 3 of those are weak.

And inside those 1,205 answers with 5 links, question-style searches do get more writing, 188 words against 173 for everything else. Real, and 15 words is not much of a payoff for the theory.

So I am calling this one open. Something decides the length and it is mostly not the number of sources, and nothing I collected pins it down. The next step is obvious enough: take the long answers and the short ones at the same link count and go read the searches.

Finding — 7 of every 8 sources already rank on that same page

Share of 39,476 links by where the page they point to sat in the organic results of the same search.

1 thing to be clear about before that number, because it changes how you read it. I collected 20 results per search and no more, and on 93.4% of searches Google returned fewer than 20 anyway, 18.7 on average. The bottom row of that chart means a site outside the 20 results I have, and some of it is sitting at 21, 30 or 50, where I did not look. 86.5% is a floor, and the real share coming out of the rankings can only be higher.

Site level, 86.5% of everything the answer links comes from a site already ranking in the top 20 right underneath it. 69.8% of it from the top 10 and 16.6% from positions 11 to 20. The AI block builds its answer out of the search results. Shocker. Anyone selling AI optimization as a replacement for rankings is selling the wrong thing.

Now split that 86.5% in 2, because the halves mean different things to different people. In 73.9% of links the answer points at the very page that ranks. In the other 12.6% the site ranks and the answer points at a different page of it.

Take that second group. 49.0% of answers contain at least 1 of these swaps. And the sites it happens to aren’t marginal. Their best ranking page sits at a median position of 4, and 83.4% of the time it is in the top 10.

What that looks like in practice. Your product page ranks 4th. The answer above it links your comparison article instead, which ranks nowhere. The traffic report shows a page you never optimized picking up visibility, and the page you did optimize looks like it lost. Nothing broke. The ranking got the site into the answer, and the model picked which of its pages to link.

A site owner and a page owner read the same data differently. Site level, a strong position is worth 86.5%. Page level, the page that earned the position is the one linked 73.9% of the time. Every number in the next section is page level for that reason.

Plain version: rank on the page and you are in the running. Do not rank and, most of the time, you are not in the answer either. That is the whole of it.

Finding — how many pages the answer actually links, rank by rank

Page level throughout this section: the address that ranks has to be the address the answer links.

Number of ranking pages linked by their own address at each position. Positions 1 to 10 in accent, 11 to 19 muted.

The gap between those 2 columns is the page swap from earlier, laid out by position. At position 1 the site gets into the answer 57.5% of the time and the ranking page itself 48.7%, a gap of 8.8 points. At position 19 it is 11.4% against 7.7%, a gap of 3.7.

The swap is something the top of the page gets. Rank near the top and the answer picks whichever of your pages it likes. Rank at 19 and it takes the page that earned the position, or it takes nothing. Everything else on this page counts the exact address, the stricter of the 2.

Across the whole study, the count runs like this. 125,104 pages ranked and the answers linked 29,172 of them by their own address, walking past 95,932. The top 10 supplied 22,755 quoted pages out of 67,506. Positions 11 to 20 supplied another 6,417 out of 57,598.

So, position 1 gets quoted 3,337 times. Position 10 gets quoted 1,261 times. Same page, same search, and rank alone cuts the count by 62%.

Turn that around and you get the line I take into planning meetings: move a page from position 10 to position 3 and its chance of landing in the answer more than doubles, from 18.9% to 42.4%. Same ranking work you were already funding, now paying in 2 places.

And here is the finding in the plainest words I have for it. The higher you rank in normal search, the more often the AI answer quotes you. That’s it. That’s the whole study. Everything else on this page is detail under that 1 sentence.

It drops off steadily, and it never gets to zero. Even at position 19, 355 pages made it into the answer sitting above them. Page 2 is worth much less than page 1, and it is still worth something. A page at 19 still gets quoted 7.7% of the time.

Finding — what the quoted pages have that the ignored ones do not

Page level again. 95,932 pages ranked in the top 20 of a search whose answer quoted somebody else. So what did the 29,172 winners have that they didn’t?

Well, not a better title, as it turns out. Sure, I checked anyway: title length, address depth, snippet length, page structure. All 4 come back the same on both sides.

How often a ranking page is quoted, by how many of the 7,601 searches its site appears on anywhere in the top 20.

What wins is how widely the site shows up. A quoted page belongs to a site that ranks for roughly twice as many searches across the market, 2,956 against 1,694. Down in positions 11 to 20 the gap widens to 3.3 times, 3,539 against 1,077.

As a rate: pages on sites ranking for fewer than 100 of these searches get quoted 9.9% of the time. Pages on sites ranking for 3,000 or more get quoted 32.0% of the time.

Take 2 pages down in positions 11 to 20. The first belongs to a site that shows up for 80 searches in this market, the second to a site that shows up for 3,500. Same stretch of the results, same kind of page, same day.

The narrow one gets quoted 4.9% of the time. The wide one 19.8%. The model picks the site it sees most often across the market. And it does that hardest down on page 2, where you would think it would matter least, because nobody scrolls that far anyway. Yeah, that one surprised me too.

About that ceiling. In the top 3, a site that shows up on under 100 searches still gets quoted 22.0% of the time, and everything above 500 lands between 46.5% and 51.5%. Showing up on more searches raises the rate up to a point. Past about 500 searches, showing up on more doesn’t change it.

One thing about the page itself does matter: the answer quotes home pages 14.6% of the time against 23.8% for inner pages. It wants the page about the thing. A company front page gets picked up a bit more than half as often as the specific page behind it.

One limit on all of that. A site ranking for 3,000 searches in a market got there on links, on brand and on trust signals, and this capture holds none of those. The coverage number may be measuring those instead, and I cannot separate the 2 with the data I have.

To answer “why this page and not that one” properly I need the backlink profile of each page, the strength of the sites linking to it, and the experience and trust markers Google leans on. Separate study, separate data. That is the next one I run.

Finding — the authority score tracks citations, and I can’t show it working on its own

So, quoted pages carry a median authority score of 84 against 80 for ignored ones. Looks like something you could work on, right? Well, look at the whole set first.

2 things in that table. First, this market is full of strong sites. 79.6% of every ranking page here sits on a site scoring above 70, and only 2.7% score 40 or under. Second, the quoted column climbs from 9.8% to 27.9%, and it climbs unevenly: the 41 to 50 band beats the 51 to 60 band above it.

Now put it next to position, measured on exactly the same pages.

Authority spreads the rate 2.8 times, from 9.8% to 27.9%, and it wobbles on the way. Position spreads it over 4.2 times, from 46.2% to 11.1%, and it never wobbles. Same pages, same day, 1 of the 2 columns is doing much more of the work.

And the position column explains part of the rest. Strong sites sit at a median position of 9, weak ones at 13. Which, yeah, is news to nobody: strong sites rank better. That is the problem with reading the score on its own, because part of what looks like the score is really just position.

All 125,104 ranking pages at positions 1 to 20, split by how many searches the site shows up on down the rows, and by its authority score across the columns.

Pick a row and read across it. The top row barely moves as the score climbs: 10, 13, 10, 10, 9, 9. The bottom row goes 37 then 30, which is the wrong way. Inside a row the score does almost nothing, and what it does has no direction.

Now read down a column instead and it runs from 9% to 37%.

I built this on all 125,104 ranking pages, then split it into positions 1 to 10 and 11 to 20 in case I was only looking at position again. Both halves come out the same shape at different heights: flat across the score, steep down the column. The top row on positions 1 to 10 runs 21, 24, 16, 17, 14, 17, and it is just as flat.

And look at the empty cells. Not 1 page in this study belongs to a site that shows up on 3,000 or more of these searches and scores 70 or under. Those combinations do not exist here, so there is nothing to compare. In this market the sites that show up everywhere are also the strong ones. Of course they are, big sites are big.

So are there strong sites that barely show up, and what happens to them

That is the question I would ask next if I were reading this, so I went and looked. Yes, there are, and they are most of the strong sites here.

2,117 sites rank somewhere in this study. 226 of them score 86 or more. And 194 of those 226, so 85.8%, show up on fewer than 100 of the 7,601 searches. 147 of them were never quoted at all. Their median is 2 searches though, so most of that group turns up once and goes away, and a rate built on 2 pages tells you nothing.

So I cut it down to sites with at least 20 ranking pages, where the rate holds still. 337 sites have that many. The table below shows 218 of them, the ones scoring 86 or more and the ones under 70, because that contrast is the point.

Read the first and third rows against each other. Same authority score, both 86 or over. The 37 sites that show up on under 500 searches get quoted 16.6% of the time. The 11 sites on 1,500 or more get quoted 30.6%. Nearly double, on the same score.

Now read the first row against the fourth. Both groups show up on under 500 searches. One has a median score of 90, the other 56. They get quoted 16.6% and 13.9%. A 34-point difference in authority is worth 2.7 points here.

The bottom row is the interesting one, and the one I trust least. 2 sites score under 70, show up on more than 1,500 searches, and get quoted 27.9%. That is what the strong wide sites get, and almost double the strong narrow ones. 2 sites is 2 sites, so I build nothing on it.

The 119 sites scoring 71 to 85 are missing from that table, and they do not land in the middle. On under 500 searches they get 10.6%, below both groups above. On 500 to 1,500 they get 18.6%, on 1,500 or more 29.3%. Same shape as everyone else, and out of order on the score again.

And the median position column keeps doing what it did before. Wide sites sit at 9, narrow ones at 12 or 13. Part of this is presence working through position, and I can’t split the 2 apart with a single capture.

So does the score do anything, once position is held still

Everything above says the score does not sort sites inside a row. A score that fails to sort a row can still carry weight of its own, and I nearly wrote the stronger sentence instead. So I tested it properly.

Take each position on its own and compare sites scoring 60 or under against sites scoring 86 or over. The strong sites get quoted more often at every position. 19 out of 19, no exceptions. At position 1 it is 52.3% against 35.0%. At position 12, 15.2% against 7.1%. At position 19, 12.4% against 3.4%.

Put those positions on a common footing and the gap is 22.3% against 13.1%. The raw gap was 27.9% against 13.1%, so position accounts for about a third of it and the other 2 thirds is still standing.

Then a harder test, and the one I would have asked for. Take pairs on the same search, 1 site scoring 60 or under and 1 scoring 86 or over, standing within 2 positions of each other. Same query, same day, same neighborhood of the results. That gives 16,996 pairs across 5,041 searches.

The strong site is quoted 20.7% of the time, the weak one 11.8%. In the 4,461 pairs where the answer took exactly 1 of the 2, it took the stronger site 66.9% of the time. At 4,461 pairs that holds.

And now hold how widely the site shows up as well

Hold the position band and how widely the site shows up, and the score stops paying. It is worth 10 points in the top 3 among sites that barely show up anywhere, worth 2 to 4 points in the middle, and slightly negative on page 2. 2 of the 6 rows go the wrong way.

What I actually think this says

The honest version has 2 halves and they do not contradict each other.

Yes, the score is associated with being quoted. Straight off the data it runs 9.8% to 27.9%, a factor of 2.8, and the strong sites win at all 19 positions and in 2 out of every 3 head-to-head pairs. Anyone saying authority has nothing to do with AI Overview visibility is arguing with the numbers.

And no, I cannot show it working on its own. Match 2 sites on position and on how widely they show up, and the extra score is worth a couple of points with no consistent direction. What it looks like from here is an order of operations: links, brand and topical strength earn the coverage and the positions, and those are what the answer reads.

Both answers depend on which question is being asked. Is a high score associated with getting quoted? Yes, clearly. Is a high score, on its own, what the model picks on? I have no convincing evidence of that, and 2 of my 6 controlled comparisons point the other way.

Either way, the practical order does not change. Work the things that earn positions and coverage across the market, and the citations follow. Chasing the score as its own goal, on the theory that the AI answer rewards it directly, is the part I would not fund on this data.

Where that lands on a budget: links and coverage still earn their money, and they earn it through positions. A proposal that promises AI Overview visibility by raising a score, with no plan for what that score is supposed to move, is selling you the middle step and skipping the 2 that matter.

Finding — the citation is the boring half, the recommendation is the interesting one

Everything above says the same thing: rank higher and you get quoted more. Fine. But now look at what a citation is actually worth to a business.

It is a small link in a row of small links, and the block hides most of them behind a Show more control. Answers here carry 5.7 links on average, and 82.6% of them carry more than 3. On most searches half the sources are off screen until somebody clicks twice. Nobody clicks twice. Getting quoted puts you in a list most readers never open.

What actually reaches the reader is the writing above that list. The block says a company name in a sentence, and the sentence recommends it. That is the thing a buyer takes away and types into the search box later. So the question I care about is how the model picks which companies to name.

The list it links and the list it names overlap, and each one holds companies the other misses, so I counted them separately.

Somebody gets recommended on nearly every search

Before asking who, I asked how often the block names anybody at all. Reading the text the way a person reads it, where a name counts whether it is a link or not, the answer names at least 1 company in 6,822 of 6,985 answers, 97.7%. Across all 7,601 searches that is 89.8%, and the difference is just the searches with no answer on them.

163 answers name nobody. I read a sample of those. They answer in categories instead of names, and they sit on the questions where the block looks careful about pointing at a specific provider.

Whether a business can appear in AI Overviews at all has a dull answer here. Somebody appears on almost every search. What matters is who, and how many of them share the same answer.

So I counted that next.

The median answer names 4 companies and the longest one names 12. Having the answer to yourself happens in 5.5% of them. What is actually on the table is a place on a shortlist of about 4, alongside 3 competitors the same reader is looking at in the same block.

How I counted, in plain terms

Every AI answer has 2 parts: the writing, and the links. A company name can turn up in 3 places, and each one is worth something different to the business.

That middle row is the one nobody measures, so it is the row this section is about. I cut every link and every piece of clickable text out of the answer first, then looked for company names in the plain sentences left behind. A name inside a link is already a link, and counting it twice would just flatter the result.

I ran this on the 50 companies the answers link to most across the study, because a name is only checkable against a site that shows up in the data at all. The unit I count is a pair, meaning “in this answer, this company”. Link to the same company twice in 1 answer and it still counts once.

2 questions, 2 ways of reading, and every table below says which one it uses. For “does the block recommend anybody, and who”, a name counts whether it is a link or plain text, because the reader sees the word either way. That covers the 97.7% above and the company table further down. For “how much of that recommendation carries a click”, I cut every link out of the text first and count only what survives in the sentences. That covers the tally and the position table.

A problem I found in my own counting

I build the list of company names out of the data itself, out of the way the block writes those names inside its own links. On single-word names that goes wrong. Say a company is called Public. Every “public record” in a sentence then reads as a recommendation, and a company nobody mentioned collects mentions.

So I measured how each single-word name is actually written across all 6,985 answers. A real company name gets a capital letter every time: 1 of them appears 6,253 times in this study and not once in lower case. An ordinary word borrowed as a name shows up in lower case constantly. 3 names were written in lower case in more than a fifth of their appearances, and there is a clean gap under them. The next name down sits at 7.0%, and after that it is zeros.

I dropped those 3, made every remaining single-word name case-sensitive, and removed 6 weak entries besides. The large companies did not move by a single count. 1 false entry fell from 732 mentions to 20. Better to find that here than in a meeting where somebody is deciding a budget on it.

A short example

One answer in the study opens with a sentence like “the easiest route is a mainstream provider such as Company D or Company E”. Neither D nor E is a link, by the way. Both are a full-throated endorsement in the first line the reader sees, and neither gets a click out of it.

Lower down the same answer links 5 sources. 3 of those companies are also named in the writing. 2 of them appear only as a link at the bottom, so the reader sees a source and never a recommendation.

That 1 answer feeds 2 of the numbers below. It joins the 44.0% of answers that recommend a company by name and send the click somewhere else, and its named companies split 3 with a link against 2 without.

What the tally says

Row by row, because the rows do not share a denominator.

• 5 companies linked, median. Across the 6,886 answers that link at least 1 of the 50, that is how many different companies each answer links to.
• 3 companies named, median. Across the 5,548 answers that name at least 1 of them in the writing, that is how many different companies each answer names.
• 67.6% of names come with a link. These 2 rows count 1 company in 1 answer, and the study holds 16,379 of those. In 11,072 the same answer also links that company, either from its source list or from the sentence itself, so the reader is told who to consider and can get there from the block.
• 32.4% of names come with nothing. The other 5,307. The answer recommends the company by name, and the block carries no link to it anywhere, source list or sentence. The reader is told who to consider and has to go type that name into a search box to reach it, which is a visit that company gets tomorrow and never connects to this block.
• 44.0% of answers carry at least 1 of those. Same event, counted per answer instead of per name: 3,073 of the 6,985. Under a third of the names turns into 44% of the answers because an answer names several companies at once, and 1 unlinked name is enough to put the whole answer in this row. The example above splits 3 with a link against 2 without.
• 60.4% of links are labeled with the company name. Of the 32,745 links, 19,778 carry the name as their clickable text. The rest are numbered markers, so the reader sees a source and no name.

So 3,073 of the 6,985 answers, 44.0%, put a company name in front of the reader and send the click elsewhere. And close to 1 naming in 3 comes with nothing to click at all.

Who actually gets named

50 companies carry the citations and the namings worth counting in this study, and they do not carry them evenly. This is the top of that list, anonymized, with the rankings each one holds on the same searches sitting next to it.

4 things I read off that table.

The list is concentrated. The top 5 hold 47.4% of every naming in the study, the top 10 hold 68.5%. 50 companies in the set, and half the recommendations go to 5 of them.

Authority does not order it. Company A leads on a score of 86, above 2 companies scoring 91. The strongest domain in the table, at 99, comes 5th. That is the same conclusion as the authority section, reached from the other end of the data.

The same rankings return very different amounts. Read the Named per ranking column. Company A turns 81.9% of its rankings into a naming. Company F ranks on more of these searches than C, D, G, J or K and turns 28.6% of them into a naming. That is the gap I could not explain earlier in the study, showing up again on the recommendation side.

Reddit ranks more than anybody and gets named least. It sits in the top 20 on 5,617 of these searches, more than any company in the table, and converts 12.4% of that into a naming. The block quotes it as a source and does not recommend it, which makes sense. Nobody can buy anything from a thread.

1 row in there is an error I could not clear with the dictionary. The payment app is named on 13.8% of answers while ranking on 3.9% of them, because the block keeps naming it as a way to pay rather than as a place to buy. A name in a sentence cannot tell those 2 apart without reading the sentence around it. I left the row in and flagged it instead of quietly dropping it.

Then the other end of the same list, which says something sharper.

These get quoted on 2% to 9% of all answers and named in the writing on close to none of them. The block takes their facts and hands the reader somebody else’s name. For a business that earns its living from the review itself, the citation is all it gets here, and the citation is the half nobody sees.

There are 2 jobs on this page and they pay differently: a site can be the source the answer is built from and still never be the thing the answer recommends.

Where the recommendation comes from

Right, so how does the model choose who to name? Same way it chooses who to link: by where they already rank.

How the next table was built, because the counting matters here.

I counted how many times 1 of those 50 companies turns up in the top 20 of a search that had an answer. Across the study that happens 82,797 times, counted as appearances. A single search usually puts about 12 of the 50 in front of the reader at once, so the 6,985 answers carry 11.9 appearances each on average. 13 of those searches have none of the 50 ranking on them at all, and the other 6,972 carry the whole 82,797.

A company holding 2 or 3 ranking pages on the same search counts once, at its best position. Take a search where a video site holds positions 7, 16 and 17. That is 1 company appearing once, at 7. Counting it 3 times would credit the same company again at the bottom of the page for pages it already won the top with, and the 11 to 20 row would fill up with second and third pages of companies that are already at the top. Counting pages instead of companies gives 90,055 rows, 8.1% more, and that version is used nowhere here.

Then for each of the 82,797 I asked 2 questions. Did the answer name the company in its writing? Did it link to it? Then I grouped by the position it was ranking at.

Read the “named” column down. A company sitting in the top 3 gets named in the writing on a quarter of its searches. The same company at 11 to 20 gets named on 1 in 12. That is a 3-fold drop, and it happens purely on position.

The linked column drops at the same speed, 63.3% to 16.4%. Both columns move together, which means there is no separate way into the writing that skips the rankings. Same work, 2 results.

One blind spot, and it runs 1 way. All 50 companies come from the link list, because I can only check a name against a site that appears there. A company the answer recommends everywhere and never links to would be invisible here. So 44.0% is a floor, and the real gap can only be wider.

Finding — Reddit gets quoted first

Now, 1 site behaves unlike everything else here, and it isn’t a company. It is Reddit. It turns up in 1,265 answers, 18.1% of all of them, with 1,417 links. That is 8th out of 739 quoted sites, and 26 times the links of an average site.

Before anything else: it ranks better than everyone else. Not slightly. Sure, that explains a lot of it. It does not explain all of it.

Now the finding, and it sits in the last 3 rows. 18.1% of its ranking pages get quoted, slightly under the 23.3% across all ranking pages. Its rankings earn every bit of the quoting it gets. What it gets on top is position in the answer: 58.2% of its links open the answer, against a median slot of 4 for everybody else.

And what that looks like on a page: a thread and 3 company pages all rank in the top 5, and the answer opens with the thread, then works through the companies. The buyer meets strangers’ opinions first and your product second.

Only 1.5% of its links come in without a top-20 ranking, against 13.5% across all sources. The model takes the opinions already ranking and moves them to the front.

It also gets quoted more than once on the same search. 101 answers, 8.0% of the ones that use it, link 2 or more separate threads. The rest of the page pulls it in too. Where the results carry a discussion block, 35.1% of answers quote it, against 15.4% where they do not.

None of this is new to anyone who has looked at a search result in the last 2 years. What I did not expect is where it is strongest. On searches with a buying word Reddit ranks in the top 20 82.2% of the time, against 74.0% everywhere else. So it sits heaviest exactly where somebody is picking a product.

It does outrank strong sites, and I want to be exact about how often. On the 5,935 searches where both Reddit and a site scoring 85 or more rank, Reddit is above the strongest of them 33.6% of the time. Usually the strong site is still ahead, and 1 in 3 is plenty to notice.

For a business that leaves 2 options, and neither is a campaign you can run in a week. Either you are in those threads, being talked about, and you pick up the reach that comes with it. Or you accept that a Reddit thread took the position and the opening slot of the answer, and you work with what is left.

Worth keeping in perspective though. It does not sell anything. Somebody who needs the product still has to leave that thread and go find a company that has it. The thread takes the click and cannot finish the job, and that is the moment the reader goes looking for a name they recognize.

Finding — 1 block on the page switches the answer off

So, read a row across, not down. The product carousel line says: on the 500 searches that show a carousel, 408 of them carried an AI answer, which is 81.6%. On the other 7,101 searches, 6,577 carried one, which is 92.6%. Same study, 2 groups, and the carousel group comes out 11 points lower.

Well, where Google draws a live data panel, the AI answer just vanishes: 3.2% coverage against 95.7% everywhere else. Product carousels and local packs push it down a little. Discussion and video blocks come with more answers: 96.6% and 94.2% against about 91% on the searches without them.

73.2% of the 616 searches with no answer at all carry 1 of those 3 suppressing blocks, against 17.8% across the whole set. Part of any keyword list is out of reach, full stop. Better to know which part before anyone puts a budget on it.

Where that leaves a plan

Read the bottom 2 rows again. Video and discussion blocks do not push the answer away. So on those searches you get the answer and a second place to appear, under your own name. Getting into a video block means making video, which costs money and plenty of businesses will not do it. It is still the cheapest way onto a result the AI answer otherwise takes over.

Same logic on the local pack. If a business can convert anybody locally, a Google Business Profile with real reviews puts the brand on the page where the answer does not reach. That is 648 searches out of 7,601, under 10% of the set, so I would not build a quarter around it. Worth knowing on the day the AI answer pushes the 10 blue links below the fold.

What I would actually do with this

All of the above is description. Here is what I change in a plan because of it, in the order I would spend on it.

1. Keep the ranking work, and re-price what it buys

Moving a page from position 10 to position 3 takes its chance of being used in the answer from 18.9% to 42.4%, on top of the clicks it already wins. That is the same work you were funding before, and now it pays into 2 places. If anyone tells you rankings stopped mattering because of AI Overviews, this data says the opposite: 86.5% of what the answer links was already ranking on that same page.

What that means when you prioritize: a page sitting at 8 to 12 on a search with an answer is worth more attention than the same page sitting at 25, and worth more than it was worth a year ago.

2. Count what the answer says about you, alongside what it links

Most reporting on this counts citations, and the block hides half of them behind Show more. The half that reaches the buyer is the writing above that list, and the block names a company there on 97.7% of these answers, 4 companies at the median.

The number worth putting in a report is how often that company is you, and who the other 3 on the shortlist were.

So the measurement I would set up is 2 numbers, tracked separately: how often the answer links us, and how often it says our name. If the second one climbs while traffic sits still, that is brand demand being built somewhere your analytics cannot see, and it usually shows up later as people searching your name.

3. Go where the answer does not reach

Video blocks and discussion blocks come with more AI answers. So on those searches there is a second place to appear under your own name, next to a block that is otherwise summarizing your competitors. Video costs real money to make, and the local pack only helps if you can convert somebody locally. Both are worth pricing on the searches that matter to you.

4. Read the Reddit threads that outrank you

On the searches where a Reddit thread and a strong site both rank, the thread is above the strongest site a third of the time, and it opens 58.2% of the answers it appears in. You cannot optimize somebody else’s thread. You can know what it says about you, and you can be present in the conversation it is summarizing.

What I take from it

What this cannot answer yet

Everything above says who gets recommended and what those sites have in common. It does not say what to do to become one of them, and I want to be straight about the difference.

The strongest thing I can put my name on is this: the companies the answer recommends cover a wide share of the market’s searches. In this set, 10 sites out of 2,261 hold 35.9% of every ranking row there is. Sit inside that group and the answer names you from positions where nobody else gets named.

The trouble with turning that into advice is that coverage is an outcome. A site ranking on 3,000 searches in a market got there on links, on brand, on trust, on years of work. I captured a search results page. None of that is in it. So “get wider coverage” describes where those sites ended up and says nothing about how they got there.

4 things would have to sit next to this data before I would tell anyone how to get recommended.

• The link profile, page by page. I had 1 third-party score per domain. A single number for a whole site cannot tell a page that earned its links from a page that inherited them.
• Experience and expertise signals. Who wrote it, what they are qualified in, who stands behind the site. None of it is visible from a results page.
• Reputation away from the search page. Where the brand is mentioned, reviewed and argued about when Google is not involved. Some of what the answer links never showed up in the 20 results I collected, and this is 1 of the places that could be coming from.
• Coverage measured against the whole niche. My 7,601 searches are 1 slice. A site that owns a different slice completely looks narrow here, and the narrowness belongs to my list.

Until those sit in 1 table, anybody selling a route into AI Overview recommendations is selling a correlation. This study is the correlation. It is worth having and worth planning around. The mechanism is still open.

That is the study I run next. It needs a crawler and a link index.