Citation overlap between AI engines is small. As of November 9, 2026, a study of 230 answers to buying questions found that two engines named the same brands about 41% of the time, while the overlap in cited domains between any pair of engines was only 6%. ChatGPT cited product pages for more than half of its sources; Claude, Gemini and Perplexity leaned on third-party comparison and best-of articles. For a marketplace seller, that split changes how a product listing should be measured. The split also explains why marketplaces read a listing four ways: same fields, different retrieval, different conclusion.
One question, four retrieval systems
The distinction matters because most e-commerce teams still measure AI visibility with a single score, usually from a single engine or an average across platforms. An average hides exactly the information that would change the plan: a listing can be the top citation in one assistant and invisible in another, and the average will report a middling position that suggests the work is half done.
Same brands, different evidence
Engines agreeing on brands 41% of the time while agreeing on sources 6% of the time is a strange combination until you consider what each is doing. The recommendation set is shaped by the question and by broad brand salience; the evidence behind it comes from whatever each engine retrieves. Two engines can name the same three products for entirely different reasons, which means improving the listing for one does not automatically improve the other.
Why the two runs never match
The study also found that asking the same question twice produces noticeably different answers, and that one engine appends the current year to most of the searches it runs in the background. Both behaviours make single snapshots unreliable. A listing review that reports “we appear in position two” without saying when, where and how many times it was run is not a measurement.
Where each engine looks
| Engine behaviour | Observed pattern | What it rewards | What it punishes |
|---|---|---|---|
| ChatGPT source mix | Product pages for more than half of sources | Complete, well-structured listings | Thin pages with no specifications |
| Claude, Gemini, Perplexity | Third-party comparison and best-of articles | Earned coverage and reviews | Brands with no third-party footprint |
| Background query behaviour | Current year appended to many searches | Listings with fresh dates and stock status | Stale specifications and old model numbers |
| Run-to-run variance | Answers differ between identical questions | Repeat measurement over time | Single-snapshot reporting |
For a marketplace seller, the first two rows imply two separate workstreams rather than one. The listing work targets engines that read product pages directly; the earned coverage work targets engines that summarise third-party comparisons. Budgeting only for the first leaves a seller invisible in three of the four systems measured.
The third row is the one most often missed. If an engine adds the current year to its searches, a listing whose title and specifications reference last year's model reads as out of date. That is a listing maintenance issue with a direct visibility consequence, and it is cheap to fix compared with content production.
Cross-system variance in one product category
Variance is not limited to software. A separate November test of luxury gift prompts ran 24 questions three times across four configurations and produced 288 answers. The leading brand appeared in 44.4% of them, but the split by system ranged from 79.2% of OpenAI answers to 1.4% of Claude answers. One challenger drew about 74% of its appearances from a single system.
Retrieval evidence was recorded for 282 of the 288 answers, and three publishers dominated: one magazine appeared in 149 of them, a fashion title in 77 and a business publication in 62. That concentration is the useful part for sellers. When a handful of pages supply most of the evidence, earned coverage in those specific outlets is a measurable goal rather than a vague ambition.
Category growth in this channel should therefore be read per engine. A seller that gains 10 percentage points of visibility in one system and loses 5 in another has not grown, even if the average rises. The only way to see that is to keep the per-engine numbers separate in reporting.
Building a listing review that survives engine changes
Start with ten to fifteen questions per category and keep them fixed. Run them across the engines that matter, at the same time each month, and record the brand named, its position, the cited domains and any factual error. Then split the fixes: listing data problems go to the catalog team, missing third-party coverage goes to marketing.
The review should also track which listings are quoted in their original form and which are paraphrased. A listing quoted accurately is doing its job; one that is paraphrased into something wrong needs its attribute structure simplified, because the model is filling gaps the seller left open.
These findings are limited to public information available as of November 9, 2026. Both studies cover defined question sets and single dates, and their samples skew toward software and luxury gifting rather than mass-market e-commerce. What transfers is the method: measure per engine, keep the questions fixed, and treat listing data quality and third-party coverage as separate budgets.
FAQ
Why do engines cite different sources for the same question?
They retrieve differently. One may read product pages directly while others summarise comparison articles, so the evidence behind a similar recommendation can differ almost entirely.
Is one AI visibility score enough?
No. Averages hide per-engine differences. Keep separate numbers so you can see gains in one system offset by losses in another.
What listing maintenance matters most?
Keeping model years, specifications and stock status current. Engines that append the current year to searches will treat stale listings as out of date.
How much third-party coverage is enough?
Enough to appear in the pages the engines actually cite. Record which domains appear in answers, then target those outlets rather than general coverage volume.
