Methodology
This report pulls together three sources of evidence. One is ours, and the other two are published studies we didn't run.
Our AI Overview audit
We looked at 39 pages in May 2026. For each one we already knew the answer: whether Google had put it in the AI Overview panel for its query or left it out. We pulled panel membership from DataForSEO, so it's observed rather than guessed.
Then we ignored the writing completely and measured structure on every page, so headings, list items and body words.
It's a small sample, and we treat it that way throughout. It tells us the direction, but it doesn't give you a number to design to.
Wan, Wallace and Klein 2024 (ACL)
What Evidence Do Language Models Find Convincing? comes out of Berkeley. The authors built ConflictingQA, a set of contested yes or no questions like "is aspartame linked to cancer", each paired with real pages from Google search that argue each side. Then they measured which page five different language models sided with.
After that, they tested which properties of a page predicted the winner. They also rewrote passages ten different ways, one property at a time, to see what moved.
This is where the relevance score and the finding about writing style come from.
Aggarwal et al. 2024, GEO (KDD)
GEO: Generative Engine Optimization introduced GEO-bench, a benchmark of 10,000 queries drawn from nine sources across 25 domains. The authors rewrote the source pages nine different ways, regenerated the AI answer, and measured how visible each page was in it.
The results we cite come from the benchmark's 1,000 query test split, averaged over five random seeds, with gpt-3.5-turbo writing the answers. This is where the evidence for extractable units comes from.
How we score
| Number | What we measure |
|---|---|
| Chunk frequency | Headings plus list items, per 1,000 body words |
| Chunk score | TAS-B relevance between the query and the passage |
| N-gram overlap | The longest run of consecutive words shared with the query |
| Extractable units | Statistics with a named source, quotations, and claims that stand on their own |
TAS-B is a DistilBERT dual-encoder trained on MS MARCO, and it's the same relevance model Wan et al. used. It's deterministic, so the same passage and query always get the same score.
It's also a proxy, though. The retrieval systems behind ChatGPT and Claude are closed, so no public model is the one they actually use.
Why we trust these numbers enough to publish them
In May 2026 we refreshed 98 existing posts on our own site against these checks, and didn't publish anything new. Cited pages went from 12.8% of the corpus to 52%, and citations went from under 100 a day to single days above 1,000. We wrote that experiment up in full in 10x AI Visibility Refresh.
Then we ran the same three checks on a client site that started with zero citations and zero cited pages. It passed 10,000 AI citations in under three months. We're not naming the client or the industry.
Neither result isolates the three numbers, because both of them bundled several changes together. So they're the reason we took the numbers seriously, but they don't prove each one works on its own.
Finding 1: AI scores chunks, not pages
When you ask an AI assistant a question, it rewrites your question into its own search queries, runs them, and shortlists a handful of pages. So far that's pretty much ordinary search.
The next step is where it stops looking like search. The model doesn't score the page. It slides a window down it, 256 to 512 tokens at a time, which is roughly 190 to 385 words, scores each slice against the query, keeps the single best one, and throws out the rest.
So what's competing for the query is your best chunk, and every H2 section is a separate candidate competing on its own.
There's a practical reason retrieval works this way. If you embed a whole 3,000 word page, it gets averaged into something vague, so a query about one paragraph only matches it weakly. The best 400 words on that same page match it strongly.
Wan et al. did exactly this to build their dataset. They ran TAS-B across each page in 512 token windows with a 256 token stride, scored every window against the question, and kept the highest scoring one.
That splits the problem in two
Gate one is retrieval. Does your chunk get pulled into the candidate pool at all?
Gate two is synthesis. Once it's in the pool, does the model use it in the answer and credit you for it?
So relevance is what gets your chunk into the pool, and extractability is what gets it used in the answer.
The two gates fail in different ways, and they need different fixes. A page that fails gate one is never seen, however good it is, and a page that fails gate two gets read and then skipped.
Most optimization work treats these as one problem, and that's a big part of why so much of it misses.
The two gate split is our own framing. The evidence under each gate is separate, so the rest of this report takes them one at a time.
Finding 2: Chunk frequency is a benchmark, not a baseline
At gate one, the first thing to ask is whether the page produces enough chunks to compete with, before you worry about whether it's any good.
Across the 39 pages in our audit, the ones Google put in the AI Overview panel averaged 26 chunks per 1,000 words. The ones left out averaged 16.
That gap is real in this sample, but 39 pages is a small sample, and the pattern didn't settle on one number. The density that got pages into the panel moved with location, industry, the type of page and the query. So we don't treat 26, or any other figure, as a baseline to hit.
Benchmark against the pages already winning
This number is really only useful as a comparison. Pick the query you want to be cited for, in the market you actually sell in. Then count chunks per 1,000 words on the pages already getting cited for it, and compare yours against theirs.
If the cited pages are running well above you, structure is probably part of why you're missing out. If you're already at or above them, it probably isn't, and the problem is further down this list.
Same words, six times the surface area
Chunk boundaries follow structure, because headings and list items are where a chunker cuts.
A 400 word section with no headings inside it gives the chunker nowhere to cut, so it becomes one candidate. If you break the same words into subheadings and a list, it becomes six.
The content didn't change at all. The page just has more separate shots at retrieval, across more queries.
We're confident that more structure means more candidates. How many candidates you need depends on your market, which is why we treat it as a benchmark and not a target.
Finding 3: Relevance is cheap math, and 105 is the target
Having candidates in the pool doesn't mean any of them win. The second number is about one chunk against one query.
This number isn't ours. Wan et al. measured a list of page properties and checked each one against how often a page won: readability, the number of unique words, sentiment, perplexity, n-gram overlap with the question, and embedding similarity to the question.
The two that stood out were n-gram overlap and embedding similarity. In the authors' words, current models "rely heavily on the relevance of a website to the query, while largely ignoring stylistic features that humans find important such as whether a text contains scientific references or is written with a neutral tone."
Both of those are cheap to compute, which fits how retrieval has to work. A system picking between candidate chunks in milliseconds can't afford expensive analysis of everyone's pages, so it leans on fast math.
Score bands
| Relevance score | Band | Estimated win rate |
|---|---|---|
| Under 95 | Weak | 30 to 40% |
| 95 to 100 | Borderline | 40 to 55% |
| 100 to 105 | Solid | 55 to 70% |
| 105 and above | Top tier | 70 to 80% |
You want 105 or higher in the chunk that holds the answer, with an n-gram overlap of at least 3. Five or more is where the phrase match signal peaks.
Reading off the paper's plots, every 5 points of relevance roughly doubles a passage's odds of winning a head-to-head. The direction comes from the paper. But the win rates in the table, and the "roughly doubles", are our reading of their correlation plots, because the paper doesn't print those figures.
What moves the score
The query here is "is intermittent fasting effective for weight loss".
The before version says "intermittent fasting" once, then switches to "this approach". That one swap costs it the phrase match and pulls the meaning away from the query. It scores 96, with an n-gram overlap of 1.
The after version repeats the term, answers in the first four words, names its source, and keeps the hedge in a subclause. It scores 107, with an n-gram overlap of 5. The facts are the same, and it's just as honest.
Here's why that works. The query and the passage are encoded separately, with no attention between them, and relevance is the dot product of the two vectors. So the passage has to carry the query's meaning on its own.
Repeating the exact term pulls the passage's vector toward the query, and a synonym pulls it away.
The easiest win is the heading. If one of the fan-out queries is literally "how does an employer of record work" and that's your H2, you've got the phrase match before the first sentence even starts. Prefixing a paragraph with the question it answers was one of the rewrites that raised win rate in Wan et al.'s tests.
Finding 4: The model can only cite what it can attribute
Everything so far is gate one, and gate two is a different job.
The model has to say something specific and be able to stand behind it. If the chunk it pulled from your page has nothing it can attribute, it'll use someone else's.
The third number is extractable units, meaning a statistic with a named source, a quotation, or a claim that stands on its own. You want at least one in the chunk that holds your answer, and more is better.
Both passages above are about remote work. The first says it "has changed dramatically since the pandemic" and that "many companies now offer flexible arrangements". None of that can be sourced, checked or quoted.
The second names Stanford's American Time Use Survey, gives 28% of full work days performed from home, up from 7% in 2019, and adds a Gallup figure. That's four things a model can quote and credit.
If both chunks scored the same on relevance, we'd expect the model to use the second one.
The evidence
This is the best evidenced of the three numbers. It comes from the GEO benchmark, where visibility was measured as position-adjusted word count, which is basically how much of the generated answer cites the page, weighted toward the top of the answer.
| Method | Position-adjusted word count | Relative lift |
|---|---|---|
| No optimization | 19.3 | baseline |
| Quotation addition | 27.2 | +40.9% |
| Statistics addition | 25.2 | +30.6% |
| Cite sources | 24.6 | +27.5% |
| Keyword stuffing | 17.7 | -8.3% |
Adding quotations beat every other method tested. Keyword stuffing was the only one of the nine that went backwards, which is a useful check if you're reading this as old SEO with a new name. What wins is being specific, with numbers, quotes and named sources.
The scores come from the paper's Table 1, which prints absolute values. The relative lifts are ours, computed against its no-optimization baseline of 19.3. The authors report the same headline themselves: the best methods improve on the baseline by about 41%.
This holds up because GEO measured attribution rather than word matching. The model can paraphrase you all it likes, as long as it can attribute what it says.
A claim with a number and a named source is a complete statement you can check, so it survives being paraphrased. "Many companies now offer flexible arrangements" can't be attributed to anyone, so it gets dropped however relevant the page was.
Finding 5: Writing for style doesn't help retrieval, and can hurt it
This is the finding we least wanted to be true.
Wan et al. rewrote passages ten different ways, changing one property at a time, and re-ran every model. Four of the rewrites targeted relevance: rewriting for relevance, adding keywords from the question, prefixing the paragraph with the question it answers, and adding a single sentence that makes the page's stance obvious.
The other six only changed style: adding a science reference, adding more information, adding contact details, and rewriting for confidence, technical language or an objective tone.
The relevance rewrites raised win rate. The style rewrites, in the paper's words, "tend to only have a neutral to negative effect on win-rate."
That runs against how most of us were taught to write. Balanced, measured and hedged is what an editor would call better writing.
"Research has produced varied results" is a true sentence, but it has no stance, no number and no entity in it. So it's weak on relevance and gives the model nothing to extract.
This is where writing for a person and writing for retrieval split apart. We don't think the answer is to write badly. You just need to be specific, so keep the hedge, but put it in a subclause after the answer instead of in place of it.
There's a tension here we should point out. Adding a science reference is a style change in Wan et al.'s setup and it didn't help, while citing sources did help in GEO. But the two studies measure different things.
Wan et al. measured which page a model believes when two pages disagree. GEO measured how much of the answer credits a page. So a reference can leave a model's verdict unchanged and still give it something to attribute.
Keyword stuffing goes the other way. Wan et al. grouped it with the relevance changes, which as a group raised win rate. But in GEO it was the only method that lowered visibility.
That fits the two gates, because repeating the query's words can help a chunk get picked, but it doesn't give the model anything new to credit.
Wan et al. publish these results as a chart rather than a table, so we're reporting the direction and not quoting a number.
What we tested and dropped
The third number used to be something else.
We originally framed gate two as verbatim lifting. AI Overview synthesis looked closer to assembling than writing, with runs of 5 to 11 contiguous words taken straight from source pages. On a sample of 41 AI Overview sentences across 9 queries, we treated 8 or more contiguous matching words as contributing to the answer, and fewer than 3 as absent from it.
One case from that sample stuck with us. For "what is customer lifetime value", all four AI Overview sentences best-matched NetSuite's page, and NetSuite wasn't in the citation list. So a page can feed the answer and get none of the credit.
We dropped it as one of the three for three reasons:
- You can only see it after the model has answered.
- It moves with every model release, as models paraphrase more.
- You can't check it on a draft.
The three numbers we kept are all properties of your own page, which is what lets you check them before you publish.
Where these numbers break
These thresholds shift the odds, but they don't decide the outcome.
The clearest counterexample came from the same 39 page audit. Sprinklr's call center KPIs page scored 97.05 on "call center metrics", which is below our own threshold, and it sat at position 3 in the AI Overview panel.
A page can miss a bar and still win on structure or authority. And a page can clear every bar and still lose to a stronger domain.
That's also why chunk frequency comes first, because in the audit, structure did more of the work than the relevance score did.
The margins that matter are pretty small. One of our own pages scored 106 with an n-gram overlap of 3, against a competitor's 103 and 2, on the same topic and intent. That competitor gets far more Google traffic than we do, and three points and one n-gram is the whole gap.
How to run the check
Check them top to bottom, on a draft, before you publish.
- Chunk frequency. Count headings plus list items per 1,000 body words, on your page and on the pages already cited for your query in your market. Aim to match or beat them, rather than hitting a fixed number.
- Chunk score. Find the chunk that holds your answer and score it against the query. Aim for 105 or higher, with an n-gram overlap of at least 3, and 5 or more at best. Put the literal question in the H2.
- Extractable units. Make sure that chunk carries at least one statistic with a named source, a quotation, or a claim that stands on its own.
When a page isn't getting cited, the first thing to figure out is which gate it's failing. If AI can't find your chunk, fix structure and relevance. If it finds your chunk and skips it, give it something to cite.
Limitations
The chunk frequency numbers rest on 39 pages. That's a small sample, and it's why we present chunk frequency as something to benchmark rather than a threshold. Within those 39 pages, 15 chunks per 1,000 words separated the two groups best, at a precision of 0.70, and a stricter rule hit 100% precision while recall collapsed.
Both of those describe this sample. Neither is a number to design to, because the density that worked moved with location, industry, page type and query.
The relevance score is a proxy. TAS-B is a public model standing in for closed retrieval systems. It predicts what an embedding retriever of that kind will prefer, but it doesn't tell you what ChatGPT's retriever actually computed.
Some figures are our readings, not printed results. The win rate bands and the "every 5 points roughly doubles" are read from Wan et al.'s correlation plots. The GEO lifts are relative lifts we computed from the absolute scores in its Table 1. The Wan style results are reported as a direction because the paper publishes them as a chart.
The published studies used older models. GEO's answers were written by gpt-3.5-turbo, and the models answering your buyers today are newer and may weigh these signals differently. The authors did test on Perplexity.ai as a live engine and saw visibility gains of up to 37%. That suggests the effect carries over, but it doesn't tell you by how much today.
The results that prompted this don't isolate the three numbers. Our own 98 post refresh and the client engagement both bundled several changes at once, with no control group. They show the approach can work, but they don't show how much each number contributes.
The client result can't be verified from here. We're reporting it without naming the client or the industry.
The thresholds are probabilistic. They come from win rate curves, not pass or fail gates. A page can clear all three and still lose, and a page can miss one and still win.
Sources: Omnipresence Research AI Overview audit, 39 pages, May 2026, with panel membership from DataForSEO. Alexander Wan, Eric Wallace and Dan Klein, "What Evidence Do Language Models Find Convincing?", ACL 2024. Pranjal Aggarwal, Vishvak Murahari, Tanmay Rajpurohit, Ashwin Kalyan, Karthik Narasimhan and Ameet Deshpande, "GEO: Generative Engine Optimization", KDD 2024. Citation data for our own site is from Bing Webmaster Tools AI visibility reporting, which covers ChatGPT and Copilot.