Ecommerce SEO Is Not AI Visibility: What 1,458 AI Shopping Answers and 60 Agent Purchases Tell Us

Summary

Ranking on Google and being recommended by AI are two different things, and this research measures the gap.

We asked ChatGPT, Gemini and Perplexity 162 real shopping questions in beauty and apparel, three times each, across three markets, and compared every answer with Google’s top 10 for the same question. Then we gave AI shopping agents 35 purchase briefs and recorded where they succeeded and where they broke. Then we changed three things on one real store and measured again 12 days later.

The headline: half of the stores on page one of Google were never named or linked in any of the nine AI answers to the same question. Google’s number one result was missing 21% of the time. Below rank three, a page-one store has roughly even odds of being named. When the agents tried to buy, the store stopped them more often than the agent stopped itself.

What follows is the data, the method and the limits, so you can check it and run it on your own store.

Why this matters for ecommerce now

By September 2026, 43% of European shoppers had already used an AI tool such as ChatGPT, Gemini or Perplexity to decide where to buy, and 18.75% said it directly influenced their final decision (Sendcloud, 8,000 shoppers, 8 markets). AI-referred traffic to US retail sites has grown 1,219% since October 2024 and converts 60% better than other traffic (Adobe Digital Insights, July 2026 data). In-chat checkout is live in the US on Google’s AI Mode and Gemini, and on Perplexity; OpenAI moved purchases out of ChatGPT and into merchant sites in March 2026, and Anthropic’s open-source Claude Commerce Agents keep checkout with the merchant.

None of that tells a store owner whether their store shows up when a shopper asks an AI, or whether an AI agent can get through their checkout. So we tested it.

What we tested

A store counted as visible for a question if it was named in the text or linked as a source in any of the nine answers. That is the most generous definition we could use.

Half of Google’s page one does not exist to AI

Of the 1,618 Google top-10 results across our 162 keywords, 50% were never named or linked in any of the nine AI answers to the same question. Counting only linked sources, not mentions, it rises to 56%.

The top three positions carry over. From rank four down, the page you fought for is roughly a coin flip across nine answers, and under one in five in any single ChatGPT answer.

This is where ecommerce SEO and AI visibility split. Ranking is a page-level contest; AI recommendation is an entity-level one, and the models only borrow Google’s judgement at the very top.

Rank matters least where purchase intent is highest

Category questions, the ones that map to keywords, still track Google reasonably well. Problem-led and comparison questions do not. The models start from the shopper’s problem and pull in stores the shopper never named.

The same pattern in three markets, worst in the smallest

Counting store-type domains only (excluding Reddit, YouTube, media and the like from the Google results): Baltic stores 53% invisible, German and Polish stores 45%, US stores 41%. The gap is real in every market. It is widest where the models have read the least.

Local presence helps, a bit. This is a correlation: local stores may also simply be more relevant to local questions.

We ran the same method on a US jewellery store later in the study (Finding 11). The curve came out the same: 48% of Google’s US top-10 results named or linked, Google #1 missing 30% of the time. The shape held across two unrelated markets and two verticals.

The three models do not agree on what Google is worth

Perplexity behaves like a search engine with a summary on top: 14 sources per answer, and it links to Google’s top 10 in 94% of answers. Your SEO still buys you Perplexity. ChatGPT and Gemini build their own shortlist. Gemini names the Google #1 result in 49% of answers but links to it in only 14%; it talks about you without sending the click. ChatGPT is the model most likely to link to your own site (26%) and the least stable on repeat asks (45%).

So there is no single “AI” to optimize for. Each model draws on its own mix of sources.

Where each model gets its information

Reddit is the most cited domain overall (335 citations); Trustpilot is third across the full study (104). ChatGPT reads stores. Perplexity reads Reddit and media. Gemini reads both.

Who the models actually recommend, and why it is not who ranks

Store mentions across the 486 answers for the Baltic market, as a worked example of what a recommendation leaderboard looks like:

Three things generalise beyond this market. A shopping mall with no online store is named as often as Zalando (30 of its 61 mentions came in answers to questions that never asked for a physical shop). A retailer from another country, Care to Beauty in Portugal, is the seventh most recommended store, almost entirely because one model, Perplexity, has read a lot about it (46 of 56 mentions). And a cross-border marketplace, ubuy.lt, was cited as a source 41 times. The models recommend what they have read about, not what is best or what ranks.

Brand mentions point the same way. In beauty, La Roche-Posay (61 mentions), CeraVe (40) and Madara (37) led; in apparel, Nike (70), Adidas (48), New Balance (37) and Audimas (33). Madara and Audimas are local brands, and both were named first more often than the global brands ranked above them. Smaller brands can win this when the models have something consistent to read.

Ask three times, get three answers

Average overlap of the full brand set between runs was 0.65: two-thirds of the brands in one answer appear in the next answer to the same question. The other third changes. Specific questions get stable answers. Open, problem-led questions (29%) are close to a lottery.

For measurement this means one screenshot proves nothing. A visibility number is a share of answers across repeated asks, per model, or it is noise.

Ranking #1 and never being named

Google #1 results in the Baltic set that got zero mentions in nine AI answers:

Ksisters.lt is the instructive one. It ranks #1 for the routine question and is never named for it, while being named five times for a different Korean-skincare question. Visibility is decided per question, not per store, which is why a keyword-level view of AI visibility is the right unit of measurement.

Another example: “Where can I buy Korean skincare in Vilnius?” Google ranks bareskin.lt, a Vilnius K-beauty shop, first. Across nine AI answers it was named once and linked never. Simitri.lt, ranked seventh, was named seven times.

When AI agents try to buy, the store is the obstacle

We wrote 35 purchase briefs with real constraints (“black leather jacket, men’s L, under €200, delivered to Vilnius this week”; “fragrance-free SPF50 under €25, delivered within 3 days”) and ran each in ChatGPT Agent mode, Perplexity’s Comet agent and Google AI Mode. Every session was stopped before payment by design. Orders placed: zero.

Google AI Mode refused all 35, including the 11 US briefs; its checkout appears only on eligible US product listings, and none matched. Perplexity’s agent credits covered 25 briefs. That leaves 60 live agent sessions.

27 of 60 sessions got the right item into a cart or to checkout. Of the 33 that did not, 24 were stopped by the store: bot walls, forced logins, site errors and stock. The agents themselves gave up or refused in only 7.

Agents with a local delivery address chose local stores 84% of the time. But only 59% of their picks were in Google’s top 10 for that brief. Delivery beats rank. Germany was the surprise: two of nine, because Sephora.pl, Douglas.pl, Allegro and Morele all blocked the agent browser.

Same brief, same agent, two cities: the black leather jacket brief run by ChatGPT’s agent.

What the AIs got wrong, checked against the live page

The team checked every price and stock claim against the store page during the session.

Eleven wrong claims in 70 sessions: Google AI Mode 6 of 35, ChatGPT Agent 5 of 35, Perplexity 0 of 25. Every one was a price, stock or size the store had not made machine-readable. The agent guessed, in the store’s name. Adobe found that LLMs can read only 61% of a typical retailer’s site (July 2026), which points to the same problem.

A 12-day experiment on one store

To test whether the gap can be moved, we took a US ecommerce client, MyRivaaz, an Indian jewellery store selling to the US diaspora, and ran the same method: 30 prompts, three models, three runs, plus the 30 Google US SERPs.

The store ranks in Google’s top 10 for 6 of 30 keywords and was named in 9.5% of unbranded AI answers, 0% on problem-led questions. Its closest competitor, Tarinika, was named in 74 of 270 answers; Etsy in 106.

We then made three changes and nothing else:

What the rerun shows: mentions rose from 20 to 28 out of 243. Real, but a two-proportion test gives z = 1.2, not significant on a sample this size. Perplexity went the other way. The structured data fix produced nothing measurable in 12 days; the site was cited 41 times before and 41 after. The entity description was partly picked up (the models mention the store’s state more often) but never in the new wording. The earned mentions line up with where the wins came: local and problem-led questions where the brand had zero before.

The answer set drifts on its own. That is the strongest argument in this report for measuring monthly rather than once.

The three-gate framework

An AI shopping agent has to get through three gates before your store makes a sale. Fail any one of them and there is no sale. Tick off what your store already does; your progress stays in this browser.

Get your store measured

Do it yourself: the prompt templates and the three-gate audit checklist used in this study are free to download. No email required.

Method and limits

Collection: consumer interfaces of ChatGPT, Gemini and Perplexity via BrightData’s AI search scrapers, logged out, September 16, 2026, one collection day; three runs per prompt per model. Google top 10 via SE Ranking’s SERP data, country-targeted to Lithuania, Germany and the United States. Agent sessions were run by hand between September 17 and 22 with VPN locations matched to each brief’s market and logged per session. The case study used DataForSEO’s interface scrapers for ChatGPT and Gemini and its API for Perplexity, so its collection method differs from the main study.

Coding: brand and store names were matched against a dictionary of about 300 names plus the domain of every cited source; a 100-row manual spot check found no false positives and about 40 missed names, which were added before the final pass. Visible means named in the answer text or linked as a source. A small number of ambiguous matches move the headline by under one percentage point.

Limits: one collection date; answers vary by account, location and time. The case study is 30 keywords and 12 days, and its result is not statistically significant. The local-domain effect is a correlation. Agent outcomes are one attempt per brief per platform. Every session was stopped before payment, so “reached checkout” means the correct item in a cart or at checkout, not an order.

The prompt templates and the three-gate audit checklist are available as a free download above. The coded dataset and agent-session log are not published.

Sources

• Sendcloud, E-commerce Delivery Compass 2026 (8,000 shoppers, 8 European markets)
• Adobe Digital Insights, AI-referred traffic to US retail sites, July 2026 data
• Ahrefs, AI Overviews and click-through rate, 300,000 keywords, December 2025 data
• Seer Interactive, AI Overviews and organic and paid CTR, September 2025
• Posti, Future of Logistics Report, February 2026 (n = 2,375)
• Eurostat, E-commerce statistics for individuals and enterprises, 2024 data
• Product.ai, Trust in AI Commerce Report, April 2026 (n = 1,463)
• Adobe, consumer survey on AI assistants in shopping, July 2026 (n = 5,000+)
• Muck Rack, What Is AI Reading?, May 2026 (25M+ cited links)
• Google, Universal Commerce Protocol announcements, January and February 2026
• The Information and CNBC on OpenAI’s Instant Checkout changes, March 2026
• Anthropic, Claude Commerce Agents, September 2, 2026
• Bain, December 2025 and Morgan Stanley, December 2025, agentic commerce forecasts

FAQ

Navneet Kaushal is the founder and CEO of PageTraffic, an SEO agency serving ecommerce and lead-generation brands across 36 countries since 2002. Request a free AI visibility snapshot or talk to us about ecommerce SEO and AI visibility.