Article contents
On a “best X” question, the engine cites the list, not the candidates
When the question is a choice, the source selected is almost always a page that compares, and almost never a candidate's page. The reason is structural: a brand site cannot answer “which one should I choose?”, because it is an interested party, whereas a comparison site supplies exactly what the model has to produce, a shortlist, criteria and an opinion. The engine retrieves that list, cross-checks it against one or two others, and writes. The brands cited in the answer are those that appear in the retrieved lists; the others do not exist in the decision.
This mechanism falls under the fourth condition for a citation, corroboration, described in our method article: on a question of choice, your claim (“we are the best for …”) is an opinion, while a comparison site's is information. A practical consequence follows, which many companies discover late: improving your own site changes almost nothing about your presence in “best X” answers as long as you are absent from the third-party pages the engine reads.
These pages are few. For a given question, in a given engine, the same five to ten domains come back from one run to the next, with variations. That is what makes the lever workable: the aim is not to appear on every comparison site in the sector, but on those the engines actually cite, and those can be identified.
Identifying the pages that decide: the survey method
The pages that decide are identified by running a panel of choice questions through each engine, several times, and counting the domains cited. The method is that of citation share, applied to sources rather than brands.
- 01Build the choice panelThirty to sixty “best”, “comparison”, “alternative to”, “X or Y” questions, per language and per area.
- 02Run several timesEach question, several times, in ChatGPT, Perplexity, Gemini (AI Overviews and AI Mode) and Claude; the answers vary.
- 03Extract the sourcesRecord each URL cited, its domain, its type, and whether you appear on it. This is the raw material for everything else.
- 04Rank by frequencyDomains ranked by number of appearances; the top ten carry most of the answers.
- 05Qualify each pageType, way in, accuracy, freshness, accessibility to AI search crawlers.
The survey reveals four types of page, which are not approached in the same way.
- 1Editorial comparisonA “ten best …” article from a media outlet or specialist blog, with criteria and opinions.Otherwise: the shortlist is drawn up without you.
- 2Sector directoryOne listing per company, filterable by city and speciality, often with reviews.Otherwise: local questions ignore you.
- 3Review platformCustomer ratings and comments, with a company profile.Otherwise: no user feedback to corroborate.
- 4Ranking or studyAnnual league table, barometer, consultancy study.Otherwise: no place in “market leaders” answers.
The relative weight of these types varies by sector and by type of question. The diagram below illustrates what we observe, for guidance, by comparing a B2B software sector with a local services sector.
Illustrative data. What to take from it: in B2B software, the editorial comparison dominates; in local services, the directory and the reviews. In both cases, brand sites are rarely the source of a recommendation.
Prioritising: the grid that stops you chasing every directory
Not every page surveyed deserves the same effort. The grid below scores each page on six criteria; you deal first with those that combine a high frequency, an absence or an error, and an accessible way in.
| Criterion | Question to ask | What raises the priority |
|---|---|---|
| Citation frequency | In how many of the panel's answers does the page appear? | A page cited in more than one answer in four for a decisive question |
| Current presence | Are you listed, and in what position? | Total absence, or a presence at the bottom of the list with no description |
| Accuracy | Is what is said about you correct and up to date? | An error on the category, the city, the price or a discontinued offer |
| Way in | Free registration, editorial selection, application, purchase? | Free registration, or editorial selection on public criteria |
| Freshness | Is the page updated, dated, maintained? | A page revised within the year, with a visible date |
| Accessibility | Can AI search crawlers read it? | Verified: otherwise the page does not count, whatever its reputation |
Table: scroll horizontally.
The frequency criterion outweighs all the others. A prestigious directory never cited for your questions has no priority; a modest comparison site cited in one answer in three does. And frequency is measured question by question: a page may decide on “best agency in Valencia” and never appear on “best agency for e-commerce”.
Getting listed: what your entry must contain
A useful entry contains the same elements as a consistent mention: exact name, category in the terms used on your site, distinguishing attribute, proof, plus the platform's own fields, area, price range, size, specialities. A half-completed listing, with an approximate name and a description copied from an old press release, is worse than an absence: it exists, but it contradicts.
- The exact name, identical to the one on the site, with no city suffix or legal form, unless the platform asks for them in a separate field.
- The platform's category closest to yours, and, in the description, your reference wording word for word.
- The factual fields all completed: city, service area, headcount, year founded, price range if the field exists. An empty field is filled in by the model with whatever it finds elsewhere, or left as an uncertainty.
- A proof: a dated figure, a named method, a public result, within the limits the platform allows.
- Recent, detailed reviews, obtained from real customers, with replies from you. Their weight in recommendations is covered in our article on customer reviews.
- A recent update date: re-read the listing at a fixed interval, at least twice a year.
For editorial comparisons, there is no listing to fill in but a writer to convince. The lever is the same as for the press: proprietary data, an expert answer, a reference description ready to be used. The writer of a comparison is looking for objective criteria to separate the candidates; supply them, with your figures and their limits, and you will be described in your own terms.
Correcting errors: the procedure
An error on a frequently cited third-party page spreads into the answers within a few weeks, and stays there as long as the page is not corrected. The most frequent case in our audits is an offer or a price discontinued long ago, still displayed on a comparison site, and repeated by the engines as if it were current. Correction follows a four-step procedure.
- Update your own site first. The correct version must exist, dated, on a page the engine can read; it is the reference you will give the publisher, and the one the engine will compare against.
- Contact the page's publisher with the precise correction, the source and the date. For a directory, the listing's form; for a comparison site, the author or the editorial team; for a review platform, the profile manager.
- Keep a register: page, error, request, date, reply, date corrected. This register lets you follow up, and link each correction to changes in the answers.
- Check in the panel, a few weeks later, that the answers now use the corrected version. If the old version persists, another source is still carrying it; you then work back up the chain.
Three things are not done. You do not ask for a competitor to be removed from a list; you do not buy reviews; you do not create your own “independent comparison” that ranks you first. These practices are spotted, and the thread that exposes them is itself retrieved by the engines. The case of errors propagated from one source to another, all the way into the engines' answers, is the subject of a dedicated article.
Paying to be listed: when it is justified
Paying for a listing is justified in one case only: the page is already cited for your questions, the listing is the only way to appear on it, and its cost is lower than what you would spend to obtain an equivalent mention elsewhere. In every other case, the platform's sales pitch, “authority”, “audience”, “AI visibility”, is no substitute for the survey: if the domain does not appear among the sources of the answers, paying will not change that.
Two precautions. A sponsored placement, flagged as such, is still a mention, but it is only worth what the listing contains; a paid and empty placement brings nothing. And the price is compared with the citation frequency, not with the position promised in the ranking: a complete listing in tenth position on a cited page is worth more than first place on a page that nobody, human or engine, consults.
Review platforms are a case apart. You do not pay to be listed on them, but for profile management features; the presence itself depends on the reviews that real customers leave. The useful investment is therefore organisational: asking for the review at the right moment, replying to each one, including the critical ones, and checking that the profile carries the same information as the site, in the same terms.
The programme over a quarter
Work on third-party pages runs in three phases, of which only the first requires tools; the other two require rigour and a register kept up to date.
- Weeks 1 to 3Survey
- Panel of choice questions, three runs per engine
- Domains cited, ranked by frequency
- Prioritisation grid completed
- Weeks 4 to 8Entry and correction
- Listings created or completed on priority directories
- Correction requests sent, register opened
- Comparison writers contacted with proprietary data
- Weeks 9 to 12Verification
- Second survey: presence, accuracy, citation share per question
- Follow-ups
- Decision on paid pages
By the end of the quarter, the list of pages that decide is stable, and maintaining it becomes a routine: one survey per quarter, a re-read of the listings twice a year, a correction as soon as an error appears. Engines change sources; a page that decided last year may have been replaced, and a new one may appear. Only the survey shows it.
What to remember
- On a question of choice, the engine cites the pages that compare, almost never the candidates' sites.
- These pages can be surveyed: a panel of choice questions, several runs per engine, domains ranked by frequency.
- Four types: editorial comparisons, sector directories, review platforms, rankings; their weight varies by sector and question.
- A useful entry carries the exact name, the site's category, all the factual fields, a proof, real reviews and a recent date.
- Correcting an error starts with your own site; you keep a register, you check in the panel, and you only pay for pages that are already cited.
Frequently asked questions
How do I know which comparison sites the engines cite for my sector?
By running thirty to sixty choice questions, the ones your buyers actually ask, through ChatGPT, Perplexity, Gemini and Claude, several times each, and recording the sources cited. The domains are then ranked by frequency of appearance. The top ten carry most of the answers; they, and not the best-known directories, form your working list.
Is there any point in appearing on a comparison site without being first?
Yes. The engine retrieves the whole list, then cross-checks it with other sources; the position on the page is not necessarily preserved in the answer. A complete entry, with your exact category and a proof, in sixth position on a cited page, is used more often than a first place accompanied by a vague or outdated description. What counts is being on the list, and being well described.
Is a little-known free directory worth the effort?
Only if it appears among the sources of the answers to your questions. Otherwise, keep the effort to a minimum: exact name, category, contact details identical to those on the site, then move on. A divergent listing on an obscure directory does more harm than an absence, because it contradicts your entity; a consistent listing costs only ten minutes and does no harm.
Can you ask a comparison site to correct a competitor's listing?
No. You only request corrections on what concerns you, with your sources. Reporting a competitor's listing, even for a genuine error, exposes you to being identified as an interested party, and editorial teams and platforms alike keep a record of such requests. Concentrate the effort on the accuracy of your own entry and on consistency with your site.
How long between a correction and its appearance in the answers?
A few weeks in general, with no guarantee: the corrected page has to be retrieved again by each engine, and then the new version has to prevail over the other sources still carrying the error. The only reliable check is the panel of questions, run again every week. If the old version persists beyond two months, look for the source that is still reproducing it.
An SEO practitioner since 2001, Kamel Malek runs SEO360 (Alicante, Valencia, Madrid, Paris). He has published three books on visibility in generative engines, including Generative Engine Optimization and Rétablir les faits.
