We Asked 5 AI Engines Which Shopping Assistant to Buy. They Barely Agreed

A vendor-selection answer inside an AI assistant is assembled from whatever sources that engine retrieved at that moment, which is why two engines asked the same buying question can return two different shortlists. In September 2026 we measured how different.

We ran three vendor-selection questions across ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode, in the United States, United Kingdom and India, and captured every source each engine drew on. The questions were the ones buyers actually ask: Best AI shopping assistant software for ecommerce, Which AI shopping assistant handles a large SKU catalogue without hallucinating?, and AI shopping assistant for Shopify stores.

The result: 324 citations across 64 distinct domains, and no consensus worth the name.

Finding 1: the five engines disagree on almost everything

Each engine's single most-cited source is different, and in four cases out of five that source barely registers with the others.

Algolia is ChatGPT's top source with 22 citations and does not appear in any other engine's top five. Webscale AI carries 13 citations on Gemini and, in this sample, all 13 come from Gemini alone. The only source with real cross-engine reach is the Shopify App Store's AI shopping assistants collection, which lands in four of the five.

For a brand shortlisting vendors, the practical consequence is blunt. Ask ChatGPT and you will hear about search-and-discovery incumbents. Ask Gemini and you will hear about the pure-plays. The two lists have almost nothing in common, and neither is wrong.

Finding 2: the answers are built from vendor-written rankings

Sorting the same 324 citations by page type explains where those shortlists come from.

Listicles are the single largest source type on buying questions. Now look at who writes them. These are the most-cited best-of pages in the sample:

Seven of the ten most-cited rankings are published by companies that appear in their own ranking. Dialog's page, for one, places Dialog first and then lists Rep AI, Gorgias, Shopify Inbox, Bloomreach Clarity, Insider One, Constructor and Luigi's Box beneath it, with each competitor's pricing quoted.

Across the whole sample, 59.9% of citations sit on vendor-owned domains against 27.5% on third-party media. In the AI-visibility category the ratio runs the other way, with independent media carrying more weight. Here, vendors write the reference material and the engines read it back.

This is not a scandal. It is a description of what the corpus contains, and the reason is unglamorous: in a category this young there is very little independent evaluation to retrieve. When no one else has done the comparison, the engine uses the comparison that exists.

Finding 3: the shortlist changes at the border

Same three questions, same week, three countries.

India's answer is dominated by enterprise search incumbents, Algolia and Coveo documentation. The UK's leans on the venture-funded pure-plays. The US sits between them, anchored on the Shopify App Store. A vendor with a UK-weighted presence is close to absent from the Indian answer to the identical question.

Finding 4: nobody owns this category

Widen the sample to all 17 non-branded buyer questions we track in the shopping-assistant category and the pattern holds at scale: 1,825 citations spread across 419 domains, with the top ten holding 21.9% between them and the single largest holding 3.4%.

For comparison, in categories with an established incumbent a single domain routinely carries double-digit share. Here the leader holds roughly one citation in thirty. Whatever consolidation is coming has not happened yet.

What this means if you are buying

Ask more than one engine. The shortlist you get is a property of the engine, not of the market. Three engines will cost you ten minutes and roughly double the names you see.

Check who wrote the ranking that named your finalist. If the page placing a vendor first is published by that vendor, it is a brochure with a table in it. That does not make the vendor wrong; it makes the ranking unusable as evidence.

Ask each finalist the five measurement questions before you believe a conversion number. We set those out in the audit of every published conversion claim in this category.

What this means if you are selling

The engines are quoting whoever published the comparison. In a category with 419 domains sharing the citations and no leader above 3.4%, the cost of becoming a cited source is lower right now than it will be again.

It also means the assistant on your own site is the second problem, not the first. An onsite assistant converts the traffic that arrives. Which brands get named in the answer that decides where a shopper goes is settled before any of that, on a different surface, by a different mechanism. Citation Rank measures your position inside those answers and Share of Voice measures your slice of them; House of Zelena moved from 22% to 56% Share of Voice on ChatGPT over six months.

The free Citation Rank scan returns your current reading without a call.

Method

Three vendor-selection prompts, run across ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode, in the US, UK and India, captured 16 September 2026. Every source each engine drew on was recorded, deduplicated by URL, and classified by page type and domain owner. Sample: 324 citations, 64 domains. The wider category figures come from the same capture across 17 non-branded prompts: 1,825 citations, 419 domains.

Two limits worth stating. Engine answers are non-deterministic, so a rerun will not reproduce these counts exactly, and the shape of the finding matters more than any single number. And a citation records what an engine drew on, not what it recommended; the two correlate but they are not the same measurement.