Can one page explain an AI shortlist? Brand coverage across 252K answers
One cited page covered the entire named set in 11.1% of measured multi-brand answers. Brand overlap does not establish which source caused the answer.
Originally published October 7, 2026
One measured cited page contained every named brand in 11.1% of eligible multi-brand answers. This tests brand-set coverage, not causal dependence on a page.
September 7–October 6, 2026 (UTC) · 251,687 measured multi-brand answers; 27,980 covered by one page
- Multi-brand answers covered
- 11.1%27,980 of 251,687 measured answers
- Measured named answers
- 259,461Includes the single-brand group
- Eligible named answers
- 323,518Before requiring measured source content
One cited page sometimes contains every brand named in an answer. In our measured multi-brand answer sample, that happened in 11.1% of answers: 27,980 out of 251,687. We call this the one-page-decides rate as a question about source coverage, not evidence that the page literally wrote or caused the answer.
The measurements come from Parse, the AI answer index Dimitry Apollonsky also runs; the analysis and interpretation are Signals'. This is an observational study, not a test of a Signals service.
The window is September 7 through October 6, 2026 on ChatGPT Search and Google AI Mode. Signals runs an aged Reddit account marketplace plus an editorial network for AI brand mentions across Reddit, Quora, Product Hunt, and Threads.
One page covered the named set in part of the sample
The headline asks whether at least one measured cited page contained every brand named in an answer with multiple brands. It returned 27,980 covered answers out of 251,687 eligible measured multi-brand answers. This is a set-overlap test: the answer's named brands must all appear among the brands detected on one of its cited pages.
The test does not require the page to contain only those brands. A broad article listing many alternatives can qualify if its detected brand set includes the answer's smaller selection. Nor does the test establish that the page supplied the names. Other cited pages and prior information can support the same answer.
The practical use is to identify a concentrated source worth reading carefully. When one page can account for the entire named set, inspect its criteria, product treatment, and limitations. That is a stronger next step than assuming each named brand must have been assembled from a separate source.
Single-brand answers are easier to cover
An answer naming one brand needs only one matching detected brand on a cited page to pass the coverage test. A multi-brand answer must satisfy a stricter condition. The figure separates answer-size groups so the easiest case cannot inflate the headline and be presented as evidence that one page explains a complex recommendation shortlist.
Single-brand answers were easier for one page to cover
Answers whose named brand set appeared on one cited page · % · September 7–October 6, 2026 (UTC)
View data table
| One brand | 76.22 |
|---|---|
| Two or three brands | 23.46 |
| Four or more brands | 8.64 |
One brand: 5,925/7,774; two or three: 9,881/42,113; four or more: 18,099/209,574. Source: Parse index, analyzed by Signals.
Each size group has its own denominator; question and source sets differ.
Each row uses its own measured-answer denominator. The groups contain different questions and source sets, so the comparison does not isolate the effect of answer length or the number of brands. A question asking for a specific product can behave differently from one asking for alternatives across a broad category.
Use the multi-brand headline when discussing the one-page idea. Keep the single-brand group as context rather than discarding it, but do not blend the groups without stating the resulting denominator. A brand-monitoring report that counts only its own name cannot reproduce this full-answer-set test without collecting the other named brands too.
The engines had different observed coverage rates
The engine comparison uses only measured answers naming multiple brands and asks the same coverage question for each engine. ChatGPT Search and Google AI Mode have separate denominators in the figure. These are observed answer populations, not matched prompts with identical sources or an experiment that restricts each engine to one selected page.
Multi-brand coverage was similar across the observed engine groups
Multi-brand answers covered by one cited page · % · September 7–October 6, 2026 (UTC)
View data table
| ChatGPT Search | 10.77 |
|---|---|
| Google AI Mode | 11.53 |
ChatGPT Search: 14,711/136,608; Google AI Mode: 13,269/115,079. Source: Parse index, analyzed by Signals.
These are observed engine populations, not matched-prompt tests.
Differences can reflect question coverage, source selection, answer length, or which brands appear on the measured pages. The result therefore does not establish that an engine relies on one page more heavily in its internal reasoning. The data expose citations and detected brand sets, not the model's sequence of reading or decision-making.
For operators, keep the engine attached to the observation. Save the full answer and its cited pages before turning an apparent source pattern into a publisher shortlist. The same page can cover the named set on one question while failing to cover another answer that names different products or includes a more specialized alternative.
Covering pages can contain substantial text
For covered answers with a positive word count, the figure reports the median length of the shortest measured cited page that covered the whole named set. Selecting the shortest qualifying page avoids automatically choosing the longest available source, while still showing why broad pages need attention when interpreting a complete brand-set overlap.
Covering pages were longer for larger named sets
Median word count of the shortest covering page per answer · words · September 7–October 6, 2026 (UTC)
View data table
| One brand | 958 |
|---|---|
| Two or three brands | 2,097 |
| Four or more brands | 3,086 |
One brand: 5,925 answers; two or three: 9,881; four or more: 18,099. Source: Parse index, analyzed by Signals.
Repeated use of the same page counts repeatedly. Only positive selected word counts enter.
The unit is words on the selected snapshot, and the observations are answers. A page used in several answers can contribute several times. Word count is not available or positive for every qualifying page, so this figure has a narrower denominator than the coverage-rate figures and should not be used to infer the length of every covering source.
A long comparison can contain many brand names without determining which ones the answer selects. Read the relevant sections rather than treating full-set overlap as proof of influence. The result does not suggest a target article length or establish that adding more products to a page increases its chance of being cited.
The denominator requires some measured source content
The analysis begins with eligible answers that name at least one brand, then identifies whether any cited page has measured detected-brand content. Only answers with that measured source coverage enter the rate figures. The raw results retain the broader named-answer count so the missing-content selection remains visible rather than being treated as a measured failure.
Even inside the measured group, some other cited pages may lack usable content. If none of the measured pages covers the whole brand set, an unmeasured page might still do so. The observed rate therefore describes the available content and cannot prove the absence of a covering source elsewhere in the answer's citations.
This boundary matters when comparing categories. A market with better crawl coverage can support a more complete set test than one dominated by inaccessible pages. We do not publish a market leaderboard from uneven small cells here. A market-level follow-up would need adequate observations and comparable content coverage before naming apparent leaders.
Latest snapshots do not reconstruct the original reading
The selected page content is the latest available snapshot before October 7, 2026. An answer earlier in the window may have encountered a different version of the page. A product could have been added or removed between those dates, so a current detected-brand set is not a verified reconstruction of the source at citation time.
That limitation applies even when the match looks exact. A page containing all named brands can be consistent with the answer without having supplied any particular name. Conversely, a page that no longer contains a named brand might have included it when the engine answered. The join cannot settle either history without a suitably timed snapshot.
For prospective work, save publication and edit receipts alongside the monitored answers. A source-content capture near the answer date would improve the comparison. The current analysis is still reproducible as a retrospective observed cut, provided its title and interpretation do not turn set overlap into a claim about the engine's internal dependence on one article.
The test does not establish exclusivity or recommendation quality
A page passes when its detected brands include the answer's whole named set, even if it names additional brands. The test does not require the engine's wording, ranking, or reasoning to match the page. It also does not distinguish favorable recommendations from negative mentions or examples used to explain why a product may be unsuitable.
Those omitted dimensions matter to a buyer. A comparison could contain the right brand names while describing different use cases, and an answer could cite it while recommending a different ordering. A name-set match alone cannot establish that earning a place on the article will produce the treatment the marketer wants.
Read the actual product discussion before deciding the page is a useful outreach target. Check whether its audience, selection criteria, and factual claims fit the brand. The brand-mentions guide explains how to turn that inspection into a concrete editorial brief without promising an outcome that the coverage test did not measure.
The second count reconciled the headline
We computed the multi-brand coverage numerator and denominator with filtered aggregates, then recomputed them using conditional sums over the measured-answer table. Both formulations returned 251,687 eligible multi-brand answers and 27,980 covered answers. This checks the aggregation while retaining the same source selection and detected-brand sets in both calculations.
The receipts map every displayed rate and word-count row to its raw output. Each rate divides covered answers by its own measured-answer group, and the headline excludes the single-brand group. The word-count figure uses only qualifying positive lengths and does not replace missing lengths with a zero-valued page.
These checks do not independently validate every detected brand. Ambiguous names, incomplete extraction, and retrospective snapshot timing remain shared limitations. A refresh should keep the counting rule and cutoff explicit, because changes in the content coverage can move the observed rate without a corresponding change in how an engine selects or names brands.
The next action is a page review, not a placement guarantee
The useful output is evidence that a particular kind of source can contain enough brand coverage to account for an answer's named set. That justifies inspecting the source closely. It does not mean the operator can purchase control of the answer, and it does not estimate the effect of adding a new brand to that page.
Start with the brand-mentions guide for the editorial brief and the page-brand naming study for the broader conditional naming baseline. The latter counts pairs, while this study counts answers and tests whole sets. Their rates answer different questions and should not be combined into one placement success estimate.
The wider Signals research collection keeps these distinctions explicit. For an actual campaign, preserve the source URL, proposed factual treatment, disclosure, and publication receipt, then observe the answers separately. That produces a reviewable result even when the source is cited but the newly included brand is not named.
Questions about the one-page-decides rate
The phrase is useful only if the underlying test remains clear: one measured cited page contains all brands named in an answer. It does not show which page the engine trusted most or how the answer was generated. The questions below preserve that boundary so a compact headline does not become an unsupported claim of editorial control.
Does one page literally decide the answer?
The data cannot show that. The study tests brand-set coverage among cited pages, not the model's internal reasoning or causal dependence on a source.
Can a page pass while listing additional brands?
Yes. The answer's named set must be contained in the page's detected set; the page can contain other brands too.
Does this identify the exact article to buy?
No. It supports closer source inspection, followed by an accurate editorial brief. It does not measure a placement service's effect or guarantee a recommendation.
How we measured
Read-only production queries with an October 7, 2026 cutoff. ChatGPT Search and Google AI Mode answers exclude gpt-5-mini and gpt-5-3-mini records while retaining blank model labels. Parse and Soar source domains are excluded. Latest source snapshots before the cutoff supply distinct detected brand sets. A measured answer has at least one cited page with detected-brand data. It is covered when one such page contains every distinct brand named in the answer; additional page brands are allowed. The headline requires at least two named brands. Repeated answers count separately. Length uses the shortest covering page per answer, retaining positive word counts. Receipts contain source SQL, verbatim results, and a number trace.
What this cannot tell us
- Snapshot content may postdate the answer and does not reconstruct what the engine read.
- Some cited pages lack measured content, even within the measured-answer denominator.
- Brand-set overlap does not establish causal influence, ordering, endorsement, or exclusivity.
- The tracked question panel is not a random sample of AI searches.
- No market leaderboard is inferred from uneven content coverage.
