Notebook · Ed. 02 · July 27, 2026
We keep hearing the same hope from founders: get your page into the model’s sources and you’ll get credited in the answer. So we ran the count. Across 15 buyer-intent queries we found roughly 112 factual claims in the synthesized answers, and only about 4 of them named their source inline. It does not work the way founders hope.
Finding 01
<4%
of factual claims named their source in the sentence that used it.
Almost never. We ran 15 buyer-intent queries through a live web-answer surface and counted 112 discrete factual claims across the 15 synthesized answers. Only about 4 of those claims, under 4%, named their source inline, meaning in the same sentence that made the claim. The other 96%-plus were paraphrased with no name attached, even though each answer listed a stack of source links at the bottom.
An inline attribution looks like “according to SparkToro, 68%.” The claim and the credit ride in the same sentence. That is the thing a reader, a journalist, or a downstream model can lift and repeat with your name still attached. By that definition, 4 out of 112 claims were attributed. Everything else was absorbed.
“Across 112 factual claims in 15 AI answers, fewer than 4% named their source in the sentence that used it. Every one that did was quoting an original statistic.”
Finding 02
~8
links pooled at the bottom of a typical answer. None tied to a sentence.
No. The link list at the end of an AI answer is a bibliography, not an attribution. Across our 15 answers we counted roughly 124 source links pooled at the bottom, about 8 per answer. That list tells you the answer read eight pages. It does not tell you which page any given sentence came from. Your pricing page can sit in that list of eight and still have its numbers restated in the answer body with no mention of your brand.
This is the distinction founders miss. Being in the retrieval pool is not the same as being credited in the answer. The eight links prove the model saw you. They do not pass your name through to the reader. The reader sees one confident paragraph, and your brand is a footnote they have to go hunting for, if they look at all.
Finding 03
4 / 4
inline credits went to a named research statistic. Every one.
Original, named, quotable statistics: a specific number with a research organization’s name fused to it. All 4 inline attributions we recorded were research stats. A Tow Center study reporting Perplexity answered incorrectly in about 37% of cases. “68% of Google searches end without a click, according to SparkToro,” based on Similarweb clickstream data (the answer also named Datos). And First Page Sage’s analysis of 29 B2B SaaS industries, estimating average CAC of $239. That was the entire list. Not one product recommendation, price, or feature comparison earned a name.
The pattern is mechanical once you see it. A number like “68%” is meaningless without knowing who measured it and how, so the model carries the source along because the source is what makes the stat usable. “Average CAC of $239” is a naked figure until “First Page Sage’s analysis of 29 industries” gives it authority. The attribution is not politeness. It is load-bearing. The number cannot stand up without it, so the two travel together.
The split
Every inline credit came from the statistical bucket. The commercial questions, where a SaaS brand most wants its name said out loud, produced zero.
| Query bucket | Queries | Claims | Inline cites |
|---|---|---|---|
| Commercial buyer-intent“best PM software”, “HubSpot cost”, “Calendly vs Cal.com” | 8 | ~62 | 0 |
| Explainer / how-to“what is llms.txt”, “how AI Overviews pick sources” | 3 | ~19 | 0 |
| Statistical / research“SaaS churn benchmark”, “zero-click rate”, “B2B CAC” | 4 | ~31 | 4 |
Method · One web-answer surface · One run · Claim counts are judgment-based, so figures are approximate
The move
Publish original data where the number is unusable without saying who measured it. The one reliable way we found to get named in an AI answer, rather than silently digested into the paraphrase, is to own a figure nobody else can produce. A benchmark. A survey. A rate computed across a defined sample. When your figure is the only source for a fact the answer wants to state, the model carries your name because it has to. Your comparison and pricing pages will keep getting absorbed. Your proprietary stat gets quoted.
The practical version: publish a number nobody else has, describe the method plainly so it reads as credible, and attach your name to it so tightly that stripping the name breaks the claim. “SaaS churn averaged X across Y companies, per our 2026 benchmark” is a sentence a model can only shorten. It cannot strip out the source without breaking the number. That is what “68% according to SparkToro” is doing, and it is why SparkToro got named in an answer on a topic thousands of pages already cover.
The honest catch
Even inside the statistical bucket, most claims still went uncredited. The churn-benchmark answer stated its numbers and named nobody. Publishing original data raises your odds of being cited; it does not guarantee it. Some of your best stats will still be paraphrased anonymously, and there is no lever that forces a model to attribute.
And this is one surface on one day. We measured a single web-answer surface, one run, 15 queries. So treat the shape of the finding as real and the exact percentages as a snapshot, not a census.
Method
We ran 15 real buyer-intent queries (8 commercial, 3 explainer, 4 statistical) through a live web-answer surface, measured here via a web-search API as a proxy for the retrieval-and-synthesis surface AI search engines draw from. This is a single surface and a single run, so read it as directional.
For each answer we recorded every discrete factual claim (a sentence asserting a checkable fact: a price, a percentage, a feature, a ranking), how many named their source inline in the same sentence, and how many links were pooled as a bibliography at the end. Deciding what counts as one claim is a judgment call, so all figures are approximate and stated with a leading “about.”
The totals: roughly 112 claims, about 124 pooled links, and 4 inline attributions, every one of them a named research statistic.
Want to know what AI says about you?
Same methodology as the audits we sell, scoped to 3 findings. Use it to decide whether to go deeper.
Filed by Jake Pereira · Founder, Citably · July 27, 2026 · Atlanta GA · ET
The next pass will widen the sample and run the same count across named engines directly (ChatGPT, Perplexity, Google AI Overviews) instead of a single proxy surface, to see whether the paraphrase default holds engine by engine.