Short answer: when several comparison pages compete for the same buying question, AI search does not pick the most authoritative one. It picks the pages whose scope matches the constraint inside the question, whose verdicts can be lifted out with a reason attached, that are visibly maintained, and whose names are confirmed by other lists. Comparison pages in AI search are cited constantly: Ahrefs found that "best X" blog lists made up 43.8% of all page types ChatGPT cited across 750 buying-stage prompts. But being cited is not the same as being believed. In our own benchmark, a vendor's list could be cited by four of five engines while the vendor itself was named by none of them.
This article is for anyone whose category gets compared: vendors whose products appear in "best X" lists, publishers who write those lists, and local businesses that turn up in "best near me" round-ups. It is about how the selection works and what follows from it. If you want the hands-on part, writing a comparison page that answer engines can use, that is a separate guide; here we stay with the question of why one list wins the citation and another does not.
Key takeaways
- Comparison and "best X" pages are the default evidence for buying questions in AI answers, across ChatGPT, Google's AI features and the other engines we measure.
- The strongest observable selection signal is scope: a question about nonprofits retrieves lists about nonprofits, not the best-known generic list.
- Freshness is visible and rewarded: most cited lists were updated recently, and in every product category we measured, more than half of the cited lists carried the year in the title.
- Self-promotional lists get cited often. Whether the author is then named depends on whether other sources agree; citation without corroboration often goes nowhere.
- Publishing many near-identical lists does not multiply your chances. Our own 34 overlapping "best tools" pages absorbed most of our impressions and returned almost no clicks.
Why comparison pages carry so much weight in AI search
A buying question is a comparison question even when it does not say so. "What should a 12-person team use for project management?" asks the engine to weigh options against a situation, and the cheapest way for an engine to do that is to find pages where someone has already done the weighing. That is what a comparison page is: a document that names several options, puts them side by side and commits to a verdict. It matches the shape of the question better than any single product page can.
The independent measurements agree on this from several directions. In Seer Interactive's tracking of 30,842 informational queries, comparison queries triggered a Google AI Overview 95.4% of the time (267 of 280), more often than any other query shape it measured, including question-format queries at 85.9%. Google itself describes AI Mode as particularly helpful for queries "where further exploration, reasoning, or complex comparisons are needed." And Ahrefs' study of ChatGPT responses across 750 top-of-the-funnel prompts in software, products and agency recommendations found that "best X" blog lists represented 43.8% of all page types among the cited sources. No other page type came close.
Our own benchmark shows the same thing from the inside. We ran 20 buyer questions per category through five models in 12 categories in September 2026 and recorded every source marked in the answers. We then sorted the cited pages by title: lists and comparisons on one side (titles with "best", "top 10", "vs", "alternatives", "compared", "ranked" or "picks"), everything else on the other. In 11 of the 12 categories, list-style pages took a larger share of citations when the question asked for a recommendation than when it did not. In electric bikes, more than four in five citations on recommendation questions went to list-style pages. In restaurants in Seattle, list pages were about half the citations on recommendation questions and roughly one in ten on the rest.
The exception is useful too. For dentists in Austin, list pages were no more prominent on recommendation questions than on others; directories, insurer pages and practice sites filled the gap. Comparison pages dominate where a category has a culture of comparison writing. Where it does not, engines fall back on whatever structured listing exists.
The chart above shows the cited domains for our project management software category. Five of the top eight cited domains belong to software vendors, and what was cited from them was mostly buying content: "best tools" lists, buyer's guides, "spreadsheet vs software" comparisons and selection checklists. That is the situation many SaaS marketers already suspect: in software categories, vendors write a large share of the comparison content, and answer engines cite it.
What happens between a "best X" question and a cited page
Nobody outside an engine can watch it choose, and any article that describes the exact ranking function is guessing. What can be described is the pipeline, because the engines document parts of it and the outputs reveal the rest. For a buying question with retrieval switched on, the path from question to cited comparison page has four observable stages.
Stage 1: the question is expanded into searches
Google says both AI Overviews and AI Mode may use a query fan-out technique, issuing multiple related searches across subtopics and data sources to develop a response. Other engines do something similar, visible in the way their cited sources cover sub-questions the user never typed. A question about "the best project tool for a creative agency managing multiple client campaigns" plausibly becomes several searches: creative agency project management software, agency resource planning, client portal tools, and so on. Each search returns its own candidates.
This is the first reason generic lists lose. The fan-out searches carry the constraint from the question, so the candidates that come back are pages written about that constraint. A famous "best project management software" list may never enter the candidate pool for the agency question at all.
Stage 2: candidates are fetched and read
The engine fetches a set of pages and extracts what it can. This is where technical access matters: a page that blocks the engine's search agent, or renders its comparison table only with client-side JavaScript, contributes nothing even if it ranks well. It is also where structure matters. A table with one fact per cell, a verdict in a sentence, and a short "best for" line per option are easy to extract. A wall of affiliate boxes with the actual reasoning hidden behind tabs is not.
Stage 3: the answer is composed from overlapping evidence
The model writes an answer that names a handful of options. When several fetched pages agree that a product is suited to the situation, that product is a safe name to include. When only one page says so, and that page was written by the product's vendor, the model has a weaker case. You cannot see the weighting, but you can see its effect: in our data, the brands named for a scoped question are usually the ones that appear across several of the scoped lists the engine cited.
Stage 4: some of the fetched pages are marked as sources
Not every page that was read is cited, and not every cited page contributed a name. Counting the retrieval pool instead of the marked sources overstated our own citation figures by two to four times before we fixed the method, which is a reminder that "the engine looked at my page" and "the engine used my page" are different events. When you study which comparison pages win, count only the sources marked in the answer body.
The same pipeline, run differently by each engine
Everything above holds across the engines we measure, but the strength of each pattern varies, and the variation changes what "being in the comparison layer" means for you. Three differences matter most.
How many sources an answer draws on.
For the same 20 project management questions, Perplexity marked 19.6 sources per answer on average, while Claude marked 5.05 and Gemini 6.35. ChatGPT and Grok sat in between at 9.2 and 9.9. An engine that cites twenty pages per answer will usually reach the long tail of scoped lists, including small independent ones. An engine that cites five has to choose, and that is where the signals described above do most of their work: with only a handful of slots, the generic list that matches nothing in particular is the first to go.
Whether the answer names anyone at all.
Not every buying question produces a list of brands. In electric bikes, ChatGPT answered 15 of our 20 questions with a guide that named no brand, and so did Claude and Perplexity. Grok and Gemini named brands more often. When an engine answers with criteria instead of names, the comparison pages it cites still shape the answer, because their criteria become the engine's criteria, but there is no shortlist to be on. For those engines, the useful question is whether your page's way of framing the decision is the one being repeated, not whether your name appears.
Whether two surfaces from one company agree.
Even within Google, the comparison layer is not one thing. Ahrefs compared Google's AI Mode and AI Overviews across 540,000 query pairs and found they cited the same URLs only 13.7% of the time, although their answers reached 86% average semantic similarity. The analysis compared single generations of each, so part of the gap is ordinary variation between runs. Either way, being the cited comparison page in one surface tells you little about the other. Ahrefs has also reported that only 37.9% of URLs cited in AI Overviews appeared in the first 10 organic result blocks, with the rest coming from deeper positions, which is consistent with fan-out searches surfacing scoped pages that the original results page never showed.
The practical consequence is that any audit of comparison pages has to be read engine by engine. A list that wins on Perplexity because it is one of twenty sources may be absent on Claude, where only five are marked. Averages across engines hide exactly the cases you need to act on.
The rest of this article walks through the four signals that are visible in the outputs. None of them is a guarantee, and none of them is a ranking factor anyone has confirmed. They are the patterns that held in our data and in the independent studies, and they are enough to explain most of the cases we looked at.
Signal one: the page's scope matches the question's constraint
The clearest pattern in our data is that engines cite comparison pages whose scope matches the situation in the question. Buyer questions almost always carry a constraint: a budget, a team type, a use case, a location, a feature that must be present. The cited lists carry the same constraint in their titles. The table below shows real examples from the project management and restaurant categories, collected on 2026-09-07.
| Constraint in the question | Titles of comparison pages the engines cited | Engines that cited a scoped list |
|---|---|---|
| A nonprofit with a limited budget | "Best Project Management Tools for Nonprofits 2026: 10 Honest …"; "12 Best Nonprofit Project Management Tools"; "10 Best Project Management Software for Nonprofits in 2026 (Free and Paid)" | All five |
| Creative agencies with multiple client campaigns | "Best Project Management Software for Creative Agencies in 2026"; "Top 16 Creative Agency Project Management Software"; "Best Creative Agency Project Management Software: Top 9 in 2026" | All five |
| Gantt charts, resource allocation and time tracking in one tool | "12 Best Gantt Chart Tools in 2026 (Tested and Ranked)"; "Best Gantt Chart Software Tools for Project Management 2026"; "Best Gantt Chart Tool for Resource Planning: 2026 Selection" | All five |
| A distributed team across three time zones | "Best PM Tools for Remote Teams 2026: 10 Tested"; "8 Best Remote Team PM Tools 2026: Built for Async Distributed Work" | Four of five |
| About 45 minutes and $20 per person in Seattle | "The Best Meals For $20 & Under In Seattle"; "Best Cheap Eats in Seattle Two People, Under $20" | Four of five |
Two things stand out. First, the cited lists are rarely the famous ones. Forbes Advisor, PCMag and similar general lists do appear in our project management data, but mostly on generic questions; on scoped questions, the narrow list written for that exact audience wins. Second, the scope match often appears in the title, not just in the body. That is consistent with the fan-out mechanism: the sub-searches carry the constraint, and titles are what search systems match first.
What this means for a page that wants to be cited
A single list that tries to serve everyone is a weak candidate for almost every specific question. The comparison pages that win are scoped: one audience, one constraint, one decision. That does not mean writing a hundred thin variants. It means choosing the handful of situations your buyers actually describe, and writing one page per situation that genuinely answers it, with different picks and different reasoning, not the same ten products reshuffled.
What it means for a brand that wants to be named
If you are the product being compared, the question is not "are we on the big list?" but "are we on the lists for the situations where we are actually the right choice?" A product built for agencies that only appears on generic lists is invisible on the agency question, which is the one it should win. Running your buyer questions with the constraint in place, and reading which scoped lists the engines cite, is the fastest way to find the lists that matter for you.
Signal two: the page is visibly maintained
Comparison content goes stale faster than almost any other content. Prices change, plans are renamed, features ship, products are acquired. Engines appear to prefer lists that show signs of maintenance, and the signs they can see are the ones a reader can see: a year in the title, a visible update date, current prices.
Ahrefs looked at 1,100 cited blog lists and found that 79.1% had last been updated in 2025, with 26% updated in the past two months alone. More of them had been updated since publication (57.1%) than had been published and left as they were. The cited lists, in other words, are mostly living documents.
Our own titles point the same way. In every one of the five product categories in our benchmark, more than half of the list pages cited for recommendation questions carried the current or previous year in the title. In local categories the share varied widely: it was low for law firms and dentists, where the cited pages were mostly directories and bar-association resources, and much higher for HVAC and real estate.
Freshness is a signal, not a trick
It is tempting to read this as "change the year in the title every January." That works only until someone checks. A list that says 2026 and quotes a price that was changed in 2025 is worse than an honest 2025 list, because the wrong price is exactly the kind of detail that ends up repeated in answers and then contradicted by the vendor's own page. Engines that read several sources will see the contradiction. Readers who click through will see it too.
What counts as maintenance is checking every fact against its source on a schedule, recording when you checked, and changing the date only when you have. The date on the page should mean "verified on", not "touched on".
Signal three: verdicts the engine can lift, with reasons attached
An answer engine does not reproduce a comparison page. It lifts parts of it: a verdict, the reason for the verdict, perhaps a price or a limit. A page whose verdicts can be lifted cleanly gives the engine something to say. A page that hedges every recommendation, or buries the reasoning in prose that only makes sense in context, gives it nothing.
Google's guidance on writing high-quality reviews describes the same properties from the reader's side. It asks reviewers to explain which option might be best for certain uses or circumstances, to discuss benefits and drawbacks based on their own research, to say why something is the best overall or best for a purpose "with first-hand supporting evidence", and to make sure ranked lists contain enough useful content to stand on their own. Each of those is also a property that makes a verdict extractable: a scope, a reason, evidence, and a self-contained statement.
Method claims are part of this. In our electric bike data, roughly a third of the list pages cited on recommendation questions announced their method in the title: "Editor-Tested Picks", "Lab Tested & Ranked", "Top 7 Tested". Whether every one of those claims is true is a separate question. The pattern suggests the claim is read, and that a list with no stated method is competing at a disadvantage against one that says how it reached its picks.
| Property | Comparison page that tends to be cited | Comparison page that tends to be passed over |
|---|---|---|
| Scope | One audience or situation, named in the title and the first paragraph | "Best X" for everyone, with the audience left to the reader |
| Verdict | "Best for agencies that bill by the hour, because time tracking and client invoicing are in the same plan" | "A great all-round option with lots of features" |
| Evidence | Stated method, prices with a checked date, named limits | Feature checkmarks with no source and no date |
| Structure | Comparison table in HTML, one fact per cell, short "best for" line per option | Facts inside images, tabs or scripts; reasoning spread across paragraphs |
| Maintenance | Visible "verified on" date that matches the facts | Year in the title, prices from an older plan |
| Candour | Says where each option falls short, including the author's own product | Every option is excellent; the author's product is first and flawless |
Signal four: the same names recur across independent lists
A single comparison page is one opinion. An engine that cites several pages for the same question has several opinions to reconcile, and the names that recur across them are the safest to repeat. This is the comparison-page version of a broader principle: a model needs the same fact from independent places before it will state it confidently.
Two observations support this. Ahrefs, looking at which brands ChatGPT recommended, reports a clear suggestion that brands positioned higher on third-party "best" lists were recommended more often, while cautioning that the comparison carries some bias and that the effect was less pronounced on long lists. And in our own data the engines converge on a small set of lists: for the nonprofit question in our project management set, the same independent list was cited by four of the five engines, and so was a vendor's list. When the engines draw on the same few comparison pages, those pages carry far more weight than their traffic would suggest, and being absent from them costs you on several engines at once.
Our data records which pages were cited, not what each page said, so we cannot check directly whether the brands named in an answer were the ones that recurred across its cited lists. The answers themselves are suggestive, though. For the nonprofit question, ClickUp, Asana and Freedcamp were each named by all five engines, while the cited lists came from at least half a dozen different publishers. A consensus that stable across engines is more likely to come from many sources agreeing than from any single page's ranking.
The vendor list that was cited but did not get its author named
The most instructive example in our data is a counter-example. On that same nonprofit question, a project management vendor's own nonprofit list was cited by four of the five engines. The vendor itself was named by none of them. The engines used the page as evidence about the category and ignored its implied recommendation of its author.
Compare that with two other questions in the same category. For creative agencies, Productive's own pages, including its list of agency project management tools, were cited by all five engines, and Productive was named by all five. For Gantt charts and resource planning, ProjectManager's own pages, including its Gantt software list, were cited by all five, and ProjectManager was named by all five. For dependencies between workstreams, ONES had its own pages cited by all five engines and was named in four of the five answers, although it is not among the most-named brands in the category overall.
The data does not say why the outcomes differ, and we will not invent a reason. It does rule out the simplest explanation. Getting your own comparison pages cited by four or five engines is not enough to get named: it worked for three vendors and did nothing for the fourth. The difference has to come from something outside the page itself, and the most plausible candidate is the one the rest of this section describes, whether other sources make the same case. That is the most practical lesson here for vendors: your own comparison page can get you into the evidence, but you should assume other lists decide whether the evidence points at you, and check them before you write another page of your own.
Self-promotional comparison pages: cited often, trusted conditionally
The uncomfortable fact about comparison pages in AI search is that many of the cited ones were written by a company that ranks itself first. Ahrefs found that across 250 "best X software"-style Google results pages, 169 (67.6%) featured a list in which the company writing the article ranked itself number one. In ChatGPT, a self-promotional blog list still showed up in more than a third of responses in the software category. Ahrefs' own conclusion is that the data points to agencies and SaaS companies publishing self-promotional "best" lists, and it immediately qualifies that the data is only one part of the equation. The author says he publishes comparison posts sparingly, avoids always ranking Ahrefs first and keeps linking to alternatives.
That qualification is where most of the risk lives. There are three separate reasons to be careful.
Reason one: citation without corroboration does not convert
As the nonprofit example shows, a self-promotional page can be cited and still fail to get its author named. If the page is the only evidence for your claim, the engine has every reason to treat it as marketing. The page then works for your competitors, because the rest of the list is useful evidence about them.
Reason two: the format is easy to abuse, and the abuse is documented
Ahrefs also found that 35% of cited best lists were published on low authority domains. Some of those are honest small publishers. Others are thin sites built to host lists. Google's site reputation abuse policy describes the related pattern of third-party pages placed on an established site mainly to exploit its ranking signals; one of its listed examples is a low-quality third-party "best casinos" page on a medical site. Editorial work, opinion pieces and properly handled affiliate links are explicitly not the problem. Lists that exist only to borrow someone else's reputation are.
Reason three: the independence claim has legal edges
In the United States, the FTC's rule on consumer reviews and testimonials makes it an unfair or deceptive practice for a business to materially misrepresent that a website it controls provides independent reviews or opinions about a category that includes its own products. A vendor's own "best X" page is not illegal. Presenting it as an independent review site is where the line sits. Disclosing that you make one of the products, on the page and near the list, removes the problem and, as far as we can see, costs nothing in citation terms: the engines cite vendor pages that say plainly who wrote them.
The workable position for a vendor is therefore narrower than "publish self-promotional lists" and broader than "never compare yourself". Compare yourself where you are genuinely a strong fit, say who you are, say where you are not the best choice, and get onto other people's lists for the same situations.
Our own case: 34 "best tools" pages and almost no clicks
We learned the volume lesson on our own site, and it was not flattering. Over the three months to mid-October 2026, insightwonder.com received 45,793 search impressions and 82 clicks in total. Thirty-four overlapping "best tools" list URLs absorbed 27,841 of those impressions, 61% of the site total, and returned 12 clicks. Across the same period, 309 queries containing "best" or "top" produced 7,909 impressions and no clicks at all.
The pages were not wrong, exactly. They were alike. Many of them compared the same vendors with slightly different framing, and they competed with one another for the same queries. A searcher who saw one of them had seen all of them. An engine that fetched two of them got the same opinion twice from the same author, which is the opposite of corroboration. And none of them was scoped tightly enough to be the obvious answer to a specific buyer's situation.
What we changed follows directly from the signals above. We consolidated the overlapping pages into fewer, distinct ones and redirected the rest, so that each remaining page answers one decision. We stopped ranking vendors, including ourselves, and replaced ranking with criteria and per-vendor facts taken from each vendor's own pricing page, with a checked date and with unknown fields left as unknown rather than guessed. Our individual comparison pages now say where a competitor is the better fit. Whether those changes improve how often engines cite us is something we are measuring, not something we can claim yet.
The lesson for anyone running a content programme
A comparison page is a bet that your page will be the one an engine uses for a particular situation. Publishing twenty variations does not place twenty bets; it splits one bet twenty ways and makes every version weaker. If your analytics show many list pages with impressions and no clicks, the fix is fewer pages with sharper scope, not more pages with new years in the title.
What to do when your category is compared in AI search
The right response depends on your position. Most organisations hold more than one of these at once: a vendor is also a publisher of its own comparison content, and a local business is both the subject of round-ups and a source of facts about itself.
| Your position | What decides whether you are named | Work worth doing | Work to avoid |
|---|---|---|---|
| Vendor whose product appears in "best X" lists | Presence on the scoped lists engines cite for the situations you fit; accuracy of your facts on those lists | Run buyer questions with constraints; read the cited lists; correct stale facts; make accurate inclusion easy by publishing plain prices and limits | Paying for placement in lists presented as independent; asking for rankings rather than accuracy |
| Vendor publishing its own comparison pages | Whether other sources agree with your page; whether the page is useful about competitors too | Few scoped pages, disclosed authorship, stated method, candid limits, checked dates | Dozens of near-identical lists; ranking yourself first everywhere; claiming independence |
| Independent publisher writing lists | Scope, method and maintenance relative to the vendor lists you compete with | Publish your method, verify facts against vendor pages, show the date you checked, write scoped lists | Year-only updates; facts copied from other lists; affiliate links without disclosure |
| Local business in "best near me" round-ups | Presence in local editorial lists and directories engines cite for your city and situation | Find the round-ups cited for your city's questions; make sure your facts there are current; earn coverage in the local lists that matter | Fabricated reviews or seeded threads; paying for "best of" badges that say nothing |
Edge cases worth checking before you act
- Categories without a comparison culture. In some local services, such as the dentists in our benchmark, nobody writes many lists, and engines rely on directories, insurer pages and practice websites instead. Writing the first good comparison page there may help, but the faster win is usually accuracy on the directories already cited.
- Questions that name your brand. "Is X good for agencies?" will be answered with your own pages and your reviews. That is reputation, not comparison visibility, and it should be measured separately; mixing it into the same score is how visibility numbers get inflated.
- Head-to-head pages written about you by competitors. If a competitor's "X vs Y" page is cited when buyers compare the two of you, check it for factual errors and ask for corrections with a link to your primary source. Responding with an equally one-sided page of your own rarely changes which page is cited.
- Lists that were accurate when written. Old lists with old facts keep being cited if nothing better exists. A plain, current page with your prices and limits gives editors and engines an easy way to update the picture.
A one-afternoon audit
- Write five to ten buyer questions without your brand name, each carrying a real constraint your buyers mention: team size, budget, use case, city.
- Run each on the engines that matter to you and save the answer text and the marked sources, not the retrieval pool.
- Sort the cited pages into comparison pages and everything else. For each comparison page, record who wrote it, its stated scope, its update date and whether you appear.
- Mark the scoped lists that cite your competitors and not you. Those are your gaps. Mark the lists where you appear with wrong facts. Those are your urgent fixes.
- Check your own comparison pages against the four signals: scope, maintenance, extractable verdicts, and corroboration elsewhere. Merge the ones that duplicate each other.
Measured this way, the comparison layer of AI search becomes a manageable list of pages rather than a black box. Some of those pages you can influence by being accurate and easy to include. Some you can write yourself, honestly scoped and disclosed. The rest you can only earn. If you want to run the question set on a schedule and see which comparison pages each engine cites for your category, InsightWonder's citation sources view does that across five engines; the audit above works by hand as well.
For the surrounding work, see third-party corroboration for auditing the outside sources that describe you, how AI assistants decide which brands to name for the broader patterns, and our project management benchmark page for the category data used above.
Sources
- Ahrefs, "Do Self-Promotional 'Best' Lists Boost ChatGPT Visibility?" (2025-12-04): ahrefs.com/blog/best-lists-research/
- Seer Interactive, AI Overview impact on Google CTR, 2026 update (2026-04-24): seerinteractive.com/insights/aio-impact-on-google-ctr-2026-update
- Ahrefs, AI Overviews vs AI Mode citation overlap: ahrefs.com/blog/ai-overviews-vs-ai-mode/; AI Overview citations and the top 10: ahrefs.com/blog/ai-overview-citations-top-10/
- Google Search Central, AI features and your website: developers.google.com/search/docs/appearance/ai-features
- Google Search Central, Write high quality reviews: developers.google.com/search/docs/specialty/ecommerce/write-high-quality-reviews
- Google Search Central, Spam policies for Google web search (site reputation abuse): developers.google.com/search/docs/essentials/spam-policies
- 16 CFR Part 465, Rule on the Use of Consumer Reviews and Testimonials: ecfr.gov/current/title-16/chapter-I/subchapter-D/part-465
- InsightWonder AI recommendation benchmark, project management software category (generated 2026-09-07; 20 questions, five models): /industries/saas-project-management/
- insightwonder.com Search Console export, last three months to 2026-10-15 (first-party).