Getting cited by an AI engine doesn't mean getting recommended

On 29 September 2026, Lightspeed Commerce published the results of a study run by the consultancy Vaer AI on how answer engines recommend retailers. Its authors put 20,000 shopping questions to ChatGPT, Gemini, Google AI Mode and Google AI Overviews between June and August 2026, across ten product categories and four North American cities. Together that amounts to roughly 460,000 answers and 2.4 million links analysed.

One finding deserves attention from marketing teams well beyond retail. Small shops and large chains appear in near-equal measure among cited sources, at about 38% each, yet large chains collect 2.5 times more top-of-answer recommendations. The AI visibility market, meanwhile, measures almost nothing but citations.

This article sets out what the study measures and what it does not. It then weighs its results against a French study published the same week, and proposes a way to track recommendations alongside citations.

What to take away

• Answer engines cite small businesses as readily as large ones, then recommend the large ones in the overwhelming majority of cases.
• Live web access narrows the gap without closing it, which suggests brand familiarity weighs as much as page quality.
• The more specific the product in the question, the smaller the share of recommendations going to small players.
• In France, the Dekuple Insight study found that the intent expressed by the buyer, rather than the sector, decides which retailers enter the comparison.
• A dashboard that tracks citations without tracking brand mentions can stay flattering long after you have dropped out of the recommendations.

What the Lightspeed and Vaer AI study measures

The protocol compares two situations: an engine answering from what it learned during training alone, and the same engine allowed to consult the web. That separation is the study's main contribution, because it isolates what live retrieval fixes from what it leaves untouched.

The authors first asked their 20,000 questions without ever using the words "local", "independent" or "big chain", to capture each engine's default behaviour. In a second round they repeated the same questions with those qualifiers added. The 200,000 answers collected without web access and the 260,000 collected with it give two comparable snapshots. Classifying retailers as large or small relied on a hand-calibrated classifier, whose verdicts a second independent model confirmed to 99.5%. The published figures also carry a 95% confidence interval (lightspeedhq.com, 29 September 2026).

The starkest result shows up in head-to-head comparisons, where the engine has to pick between two shops without consulting the web: it goes with the larger one 90% to 94% of the time. Measurements also vary from city to city, from 70-77% large chains in Los Angeles down to 57-62% in Montreal. That geographic spread is a reason not to carry the results straight over to a European market. The study covers four cities in North America, and its full technical report is available on request only.

Parity among sources does not carry through to recommendations

The gap between the source consulted and the retailer named is the heart of the matter. An engine can read an independent shop's page, use it to confirm a price or a stock level, then recommend a national chain in the body of its answer.

The study's figures show this in sober terms. Among cited sources, large chains and small shops each account for roughly 38%, an almost even split. Among recommendations placed at the top, large chains are chosen 2.5 times more often than small shops. Independent retailers do make it into the raw material of the answer, then rarely into its conclusion.

The breakdown by condition shows how this plays out:

Web retrieval therefore improves the position of small players considerably, without restoring it. Google AI Overviews stands out in the other direction, with 68% of its answers naming no local shop at all. Conversely, small shops' share of recommendations falls from roughly a third to a tenth as soon as the question names a specific product rather than a category.

Why does an engine cite a source without recommending its author?

Citations and recommendations rest on different signals: the first rewards a page that answers, the second a brand the engine recognises. The work published over the past year converges on this point, though none of it establishes strict causation.

The Ahrefs study of December 2025, covering 75,000 brands, measured a correlation of 0.656 to 0.664 between how often a brand is mentioned across the web and its visibility in AI answers. The same study found only 0.218 for backlinks. Its authors note that this is a correlation, not evidence that one more mention produces one more recommendation. The association nonetheless sits comfortably alongside what Profound observed in January 2026 across 27 million citations. On a generic question, a brand's own site accounts for just 4.3% of displayed citations. That share climbs to roughly 58% once the question contains the brand name.

A measurement published by Victorious in July 2026 across 150 to 175 brands extends the picture: 96% of brands are described accurately when named directly, while 89% fail to appear in answers to a category question. Engines can describe these businesses, then do not call on them unprompted. For a little-known company, the difficulty is not being understood but being summoned. We ran into the same asymmetry when looking at brand citations that never show up among displayed sources.

In France, buyer intent reshuffles the ranking

A French study published the same week adds a useful qualification: retailer size does not decide alone, because the intent expressed in the question redraws the competitive set each time. Dekuple Insight analysed 3,882 answers from ChatGPT and Gemini over two months, across six retail sectors, at 45 queries per sector, and identified 1,719 retailers in them.

The study's dominant retailers do not dominate uniformly. Sephora appears in more than 70% of beauty-sector answers, with an average rank of 1.59. Decathlon features in 80% of ChatGPT answers about quality in sport, yet comes first in only 30% of them. Diversity varies sharply by sector, from 496 distinct brands across 560 answers in fashion down to 113 brands across 630 answers in food (ecommercemag.fr, October 2026).

The authors' methodological observation matters to any company, whatever its size. Engines do not reason by market category but by the need expressed: price, speed, expertise or trend each surface a different list of competitors. On a price criterion in beauty, E.Leclerc and Action end up compared with specialist players. Thomas Zavrosa, chief executive of Dekuple Insight, describes these engines as "a new layer of prescription between brands and consumers". The useful question, in his view, is which intents make a brand relevant and what evidence supports its recommendation.

One last observation from this study has immediate practical reach. When the answer turns on an operational fact, engines lean first on retailers' official websites. In consumer electronics, those sites account for 91% of the domains drawn on for click-and-collect, availability or delivery times. On evaluative criteria, by contrast, engines turn to third parties: specialist media, consumer associations, price comparison sites and review platforms.

A single word moves the result

The vocabulary of the question weighs more than most of the technical settings the industry recommends. The Vaer AI study tested adding qualifiers to the original 20,000 questions, and the effects differ considerably from one word to another.

Adding "independent" to the question drops the share of sources coming from large national chains from around 44% to 9%, and more than doubles the share of small shops recommended. The words "local" and "near me", by contrast, produce only a limited effect. The asymmetry suggests that engines do hold a representation of independent retail, but activate it only on precise vocabulary.

You do not choose how your customers phrase things, so this observation is not a tactic you can apply directly. It does indicate which vocabulary your own pages should carry. A page that never describes itself as an independent shop, a workshop, a practice or a family business leaves the engine to guess which category it belongs to. The Censuswide survey conducted for Lightspeed among 2,000 North American consumers indicates demand for these businesses: 50% of respondents say they would buy closer to home more often if AI made independents easier to find. That sample is self-reported and limited to North America.

What can a company that is not a national chain do?

Three pieces of work stand out in the available data, and none of them is about technical markup. They concern the evidence you publish, the mentions you earn elsewhere, and what you choose to measure.

1. Publish the operational facts engines come to you for. The Dekuple Insight finding on official websites is the most directly actionable part of that study: opening hours, lead times, collection arrangements, service areas, return conditions. This information needs to sit in readable text on a stable page, not inside an image or a widget loaded after the fact. A page like that serves as a source even when the recommendation goes to a better-known competitor, and it puts you in the raw material of the answer.

2. Work on mentions beyond your own site rather than on content volume. The correlation Ahrefs measured between brand mentions and AI visibility remains a correlation, but it points the effort towards sector directories, review platforms, local and trade press, and third-party comparisons. The Trendos study of August 2026, covering 107 million answers, runs the same way in B2B software. A brand's own site accounts for just 2% of citations there, against 47% for review sites. That split inverts in consumer goods, which rules out treating it as a general law.

3. Measure recommendations separately from citations. This is the practical consequence of the 2.5-fold gap Vaer AI found. A two-column table is enough to start. For each question in your panel, record whether your site appears among the displayed sources, then whether your name appears in the text of the answer. Citation tracking by page and by engine covers the first column, while the second calls for reading the answer itself. The two signals move independently, and the gap between them is precisely what to watch. We set out this verify-after-the-fact logic in our Track, Fix, Prove method.

A fourth point concerns the stability of the measurement itself. AI answers vary from one run to the next, and the gaps between engines stay considerable, as the sources ChatGPT and Perplexity do not share illustrate. A single measurement on a single question establishes nothing, and a fixed panel tested regularly is the minimum needed to tell a trend from a fluctuation.

Measure what the answer recommends, not only what it cites

Dax Dasilva, founder and chief executive of Lightspeed Commerce, framed the stakes in an op-ed published on 2 October 2026: "The concern isn't simply that AI might steer shoppers toward bigger retailers. It's that AI could increasingly determine which businesses get the chance to compete at all." The line comes from an executive whose company sells tools to independent merchants, and it is worth reading with that interest in mind.

The data published in late September 2026 does not describe a closed door, though. It describes a selection that happens after the sources have been read, on familiarity signals most AI visibility dashboards never look at. A company counting its citations can believe it is well placed while its name appears in no answer at all.

If you change one thing in your tracking this year, add that second column. Record, question by question, what the answer recommends alongside what it cites. The gap between the two will tell you whether your work addresses the right problem, and whether the next push should go to your pages or to your presence elsewhere. We have set out elsewhere the GEO techniques research actually supports, and the hierarchy they describe likewise puts perceived authority ahead of markup.

Run a free AI visibility audit to see whether answer engines already read your site, and on which questions your name is still missing.