Best-of lists in AI answers: 568K cited pages and what happened to their brands
568K listicle-shaped cited pages, a naming comparison, and a sampled before-and-after cohort. First observed citation is not publication.
Originally published October 7, 2026
Listicle-shaped pages represented 20.6% of distinct cited sources. The retrospective entry comparison is not placement lift.
July 9–October 6, 2026 (UTC) · 567,704 listicle-shaped pages of 2,751,223; 6,658 paired entry observations
- Share of cited pages
- 20.6%Distinct source-page denominator
- Listicle-shaped pages
- 567,704Title/URL proxy
- Paired entry observations
- 6,658Selected page-brand cohort
Listicle-shaped pages were common sources, but inclusion was not a naming guarantee. We classified 567,704 of 2,751,223 distinct cited pages as listicle-shaped, or 20.6%. A separate page-brand sample and a restricted before/after cohort show why the share should not be presented as measured lift from buying a roundup placement.
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 page-share window is July 9 through October 6, 2026; the other figures state their own windows. Signals runs an aged Reddit account marketplace plus an editorial network for AI brand mentions across Reddit, Quora, Product Hunt, and Threads.
About a fifth of distinct cited pages matched the list pattern
The measured share is 20.6%, representing 567,704 listicle-shaped pages among 2,751,223 distinct cited pages. Each source page counts once in this comparison, regardless of how often it appeared in answers. This is a page-composition result, not the proportion of answers using a list or the probability that a published roundup gets cited.
The pattern uses best-of, top-list, and roundup wording in titles and URLs. It is deliberately described as listicle-shaped because matching language does not prove every page is an ordered product comparison. A tutorial can match, and an unlabeled comparison can be missed. The exact expression is recorded in the receipts.
For operators, the result justifies inspecting relevant comparison pages in the source set. It does not justify buying any page that resembles a list. The relevant question is still whether the article answers the buyer's need and can represent the product accurately within its editorial criteria.
Brands on listicle-shaped pages were named less often per pair
In the separate September 7 through October 6 naming sample, listicle-shaped page-brand-answer pairs had a 27.94% naming rate, compared with 39.03% for other page shapes. The denominator includes every detected brand on each measured cited page, so a broad roundup can contribute many brands that a particular answer does not select.
Listicle pairs had a lower observed naming rate
Measured pairs whose brand appeared in the answer · % · September 7–October 6, 2026 (UTC)
View data table
| Other shapes | 39.03 |
|---|---|
| Listicle-shaped | 27.94 |
Other: 1,119,792/2,868,914; Listicle: 673,798/2,411,170. Source: Parse index, analyzed by Signals.
This is a separate snapshot pair population, not a controlled format experiment.
This result should not be rewritten as an experimentally established advantage for other formats. Page groups differ in brands, topics, prominence, length, and the buying questions they answer. The shape classifier is also heuristic. The figure is a descriptive comparison that challenges an unconditional listicle-placement promise, not a causal ranking of editorial formats.
The brand-naming study reports the broader engine and prominence splits. For a brief, specify the product's relevant use case and supporting evidence rather than assuming inclusion alone will make the engine choose it. A cited list and a named brand remain different observations.
The entry comparison starts at first observed citation
The entry date is Parse's first recorded citation, not the page's publication date. Our entry cohort samples up to 1,000 listicle-shaped pages, ordered deterministically by a hash of their source identifier, first observed from June through August 2026 that also appeared in the later naming panel and had measured brand content available before the cutoff.
This later-citation requirement creates survivorship selection. Pages that disappeared from the later source panel do not enter the cohort, even if they were first cited during the entry window. The latest snapshot can also contain brands added after the entry event. Both limitations make this a retrospective association study rather than a placement experiment.
The before window covers the fourteen days preceding first observed citation; the after window covers the following thirty days. We use the recorded rolling naming rate and require brand–question combinations observed on both sides. The result is therefore a comparison of measured rolling rates on a matched observed panel.
The paired cohort has its own denominator
The paired entry analysis contains 6,658 page-brand pairs across 945 pages and 3,774 brands. We average the observed rolling naming rate within each side of the event, preserve the same brand–question combinations, and then summarize the eligible pairs. These are not all the listicle-shaped pages counted in the headline share.
Paired rolling naming rates around first observed citation
Mean recorded seven-day rolling naming rate · % · First observed citation: June–August 2026; 14 days before and 30 days after
View data table
| Before | 34.89 |
|---|---|
| After | 25.73 |
6,658 page-brand pairs; 945 pages; 3,774 brands; brand-question combinations observed on both sides. Source: Parse index, analyzed by Signals.
First observed citation is not publication. The selected surviving cohort has no untreated control.
The mean recorded rate was 34.89% before and 25.73% after. The difference is descriptive. The rolling measure overlaps across adjacent days, observation density can vary, and the event was not assigned randomly. A common brand can also appear on several pages and contribute several page-brand pairs.
Use this figure to understand what happened around an observed source-entry event in the selected cohort. Do not label the difference placement lift. We lack a matched untreated cohort and publication-level treatment evidence, and the selected pages already had to survive into the later citation panel to appear here.
Entry-month groups are different cohorts
The monthly figure groups page-brand pairs by the month of first observed citation and reports each group's before and after rates. The dates are chronological, but the points do not follow one unchanged set of pages through time. They are separate entry cohorts and should not be read as a forecast or a continuous campaign trajectory.
Entry months represent different selected cohorts
Mean recorded seven-day rolling naming rate · % · First observed citation: June–August 2026; 14 days before and 30 days after
View data table
| 2026-06 | 47.63 | 30.99 |
|---|---|---|
| 2026-07 | 30.55 | 24.98 |
| 2026-08 | 25.9 | 20.93 |
2026-06: 2,278 pairs; 2026-07: 2,223 pairs; 2026-08: 2,157 pairs. Source: Parse index, analyzed by Signals.
Points compare distinct entry cohorts, not a continuous trajectory of the same pages.
A difference between June and August can reflect the brands, topics, or sources entering the index in those months. The observation windows also encounter different engine conditions. Matching brand–question combinations within each event reduces a coverage problem, but it does not make the monthly groups equivalent or remove changes in the surrounding answer ecosystem.
The figure is most useful as a check against treating the pooled average as universal. A refresh should preserve the cohort definitions and report their sample sizes. If later cohorts lack enough follow-up, they should remain absent or unmeasured rather than being drawn as a zero-valued continuation of the line.
The study did not measure the effect of list position
We did not reconstruct product order from every article and compare it with recommendation order in the answer. The page-shape pattern identifies likely lists, but it does not determine whether a page has a ranking, uses category labels, or places the brand in a numbered position. A title match cannot answer an order-copying question.
That matters for a vendor offering a premium slot. The share of cited listicle-shaped pages is not evidence that a higher paid position changes how an engine ranks the brand. A separate text-level study would need reliable list extraction, brand matching, and comparable answer-position measurements, with ambiguous articles handled explicitly.
The current operator recommendation is editorial: ask why the product belongs, what the ordering means, and whether the treatment is accurate. The best-of-list guide explains that brief. It does not convert an unmeasured position relationship into a suggested slot or a guaranteed recommendation rank.
Publication-to-citation lag remains unmeasured
The first observed citation date cannot establish how long a page waited after publication. A page may have existed for months before entering the tracked source set. The entry analysis therefore uses an observation event, not a publication event, and its before/after movement cannot supply a waiting-time promise for a newly purchased roundup placement.
The Signals ledger does not provide the complete delivered-listicle URLs and publication receipts needed for that service cohort in this batch. We also do not infer those receipts from a page's crawl time. A crawl says when content was observed, not when the article or brand mention was first published.
For future work, save the final URL, original publication date, brand-inclusion date, and later edits. Then record the first observed citation and first observed brand naming separately on a fixed question panel. That would make a prospective lag description possible, though a causal placement-effect claim would still need a stronger comparison design.
The page-share check agrees through a different join
The headline first counts distinct source IDs within the eligible citation set, then checks the result by deduplicating answer citation arrays before joining the classified source pages. Both formulations return 2,751,223 distinct pages and 567,704 listicle-shaped pages. The agreement verifies the counting while retaining the same title-and-URL classification rule in both formulations.
The naming comparison has a different denominator and is traced to its own pair-level raw output. The entry results likewise use a selected matched cohort and the recorded rolling rate. The number trace keeps those populations separate so a reader cannot accidentally apply an entry-cohort percentage to the headline page count.
The receipts also preserve the exact windows and exclusions. Rerunning after additional crawls can change which brands are detected on a page, even for the same historical answer window. A refresh should therefore state the new content cutoff and avoid presenting retrospective snapshot changes as evidence that a placement improved over time.
The operator should brief the comparison's useful distinction
The evidence supports investigating relevant roundups, then checking the product treatment rather than buying a list label. A useful article explains which buyer should choose each option, what the product can do, and where its limitations matter. Those details make the comparison worth reading without requiring a promised AI effect that this study cannot establish.
Use the best-of-list guide for the publisher shortlist, factual brief, and delivery record. The wider Signals research collection keeps source composition separate from service outcomes. The studies are complementary evidence, not a table of interchangeable placement returns measured under one experimental design.
After publication, inspect the actual answers. Save the target page's citation and the brand's wording separately, with the question, engine, and observation date. That lets the campaign report preserve partial or absent results honestly. Inclusion on a useful page can be verified; a resulting recommendation needs its own observation and interpretation.
Questions about the listicle result
The study contains several populations because page prevalence, brand naming, and movement around first observation are different questions. Keep the relevant population attached to each quoted number. In particular, the entry cohort's result should not be presented as the experience of every cited listicle or as a measured effect of a placement service.
Does first cited mean first published?
No. It means the first citation observed by Parse. The article could have been published much earlier.
Did listicle brands get named more often than other brands?
Not in the separate pair-level comparison here: the listicle-shaped group had a lower observed naming rate. That difference is not a causal format effect.
Is the before-and-after difference placement lift?
No. The cohort is retrospective and selected, uses rolling rates, and has no matched untreated placement control.
How we measured
Read-only production queries frozen at October 7, 2026. Answers are ChatGPT Search and Google AI Mode in the stated UTC window, excluding gpt-5-mini and gpt-5-3-mini records while retaining blank model labels. Citation IDs are expanded from eligible answers. Parse and Soar source domains and subdomains are excluded. Latest content snapshots are selected strictly before October 7. SQL, raw output, and number trace are committed under docs/research-receipts for this slug. Page shape uses the title/URL regex documented in SQL. The entry cohort samples up to 1,000 pages ordered by MD5 of source ID that first appeared June through August and also had citations in the later 30-day panel. Latest measured brand content is joined to daily seven-day rolling naming rates 14 days before and 30 days after first observed citation. Only brand-question combinations with observations on both sides contribute; their within-window means are averaged per page-brand pair.
What this cannot tell us
- First observed citation is not publication; lag cannot be estimated.
- Later-panel selection creates survivorship bias; latest brand content can postdate the entry event.
- Rolling rates overlap across days; no matched untreated cohort or engine-specific entry effect was measured.
- The tracked question panel is not a random sample of all AI search.
- This observational analysis does not establish a causal effect of a Signals service.
- Entry rates use the stored cross-engine rollup; they are not refiltered by answer model.
