Every Published AI Shopping Assistant Conversion Claim, Audited

A conversion lift claim from an AI shopping assistant vendor is a comparison between two groups of shoppers, and the entire meaning of the number sits in which two groups were compared. Rep AI publishes 10-30%. Envive publishes 4x. Alhena publishes 20% AOV at Victoria Beckham. Gorgias publishes 14% of conversations ending in an order. All four can be true at once, because they are not measuring the same thing.

We collected every published claim in this category we could find on 23 September 2026, quoted each one verbatim from the vendor's own page, and worked out what each is actually counting. This is not a teardown. Two of these vendors publish more methodology than most ecommerce software companies ever do. It is an attempt to make the numbers comparable, because right now they are not.

The claims, as published

Every figure below is quoted from the vendor's own website, captured 23 September 2026.

Read that table twice and the problem announces itself. Four of the five vendors report a lift. None of them reports a sitewide conversion rate before deployment and after.

The four choices that decide the number

A conversion lift figure is the output of four methodological decisions, and changing any one of them moves the number by a multiple.

The denominator decides almost everything

Engaged-visitor conversion rate is the share of shoppers who interacted with the assistant and then ordered. Sitewide conversion rate is the share of all shoppers who ordered. The first is nearly always several times the second, and a vendor can quote it honestly while telling you almost nothing about what the tool did to your business.

Envive is explicit about this, and deserves credit for it: its 4x figure is stated as applying to "visitors engaging with the Adaptive Storefront". Rep AI's "10-30% Lift in CVR" does not say. That is the single most consequential missing word in this category.

Attribution windows are mostly unpublished

Gorgias publishes a three-day window and defines it plainly: "An order is 'influenced' when a shopper interacted with the Shopping Assistant before completing a web purchase, within the assistant's three-day attribution window." Widen that window and the influenced share rises without the product doing anything differently. Nobody else in the table publishes a window, so none of these numbers can be lined up against each other.

Gorgias goes further than anyone else here. Its own research states the method is "associative attribution, not proof of cause", and discloses that roughly 80% of influenced orders rest on the weakest available signal, a conversation that merely preceded a purchase, with about 18% carrying a stronger product-suggestion signal. That disclosure is the most useful sentence published in this category so far.

Self-selection runs the whole comparison

Shoppers who open a chat widget are not a random sample of your traffic. They are further along, more motivated, and more likely to buy before the assistant says a word. Comparing them to everyone else measures intent as much as it measures software.

Retail Systems put this to three analysts in its September 2026 piece, and Julian Skelly of Publicis Sapient drew the distinction as sharply as it can be drawn: "A 25 per cent uplift in conversion from users who engage with the AI assistant is a very different claim to a 25 per cent uplift across the full digital estate."

Nothing is ever deployed alone

Assistants go live alongside site redesigns, peak-season discounting, new email flows and paid pushes. Nick Dutton of Leading Resolutions, in the same piece: "The AI assistant was likely deployed alongside other variables, such as targeted discounting. Without transparent attribution modelling, it is difficult to credit the tool by itself."

James Heimers of RAPP adds the baseline problem: "In most cases this is a relative uplift on a low baseline, so the absolute gain is more modest." A jump from 0.8% to 1.0% is a 25% lift and two tenths of a point.

The retailer numbers are steadier than the vendor numbers

The most credible figures in this category are coming from retailers reporting their own results, not from the companies selling the software.

Frasers Group reported "conversion uplifts of up to 25 per cent compared with traditional search journeys". Kingfisher reported "significantly higher conversion rates, in some cases more than double, from customers who engage with our AI agents", alongside "around £165 million of group sales in FY25/26" and "more than 60 per cent year-on-year growth" in usage. Both are quoted in Retail Systems, September 2026.

Note what both retailers did that the vendors did not: they named the comparison group. Frasers says "compared with traditional search journeys". Kingfisher says "customers who engage with our AI agents". Those are engaged-visitor comparisons, stated as such, and they are more useful than an unqualified lift precisely because they are bounded.

The five questions to ask before you sign

Ask these in the demo. A vendor with real measurement answers all five in under a minute.

1. What is the denominator? Engaged visitors, or all sessions? If it is engaged visitors, what share of traffic engages?
2. What is the attribution window, and what happens to the number at one day and at seven?
3. What was the comparison group, and was it matched? A holdout is the only clean answer.
4. What else changed on the site during the measurement period?
5. Is the figure relative or absolute? Give me both, and give me the baseline.

If the answers do not arrive, you are not being sold a measured outcome. You are being sold a plausible one.

What we can and cannot tell you

We are not going to add a conversion number of our own to that table. Tru Commerce does not have a published onsite-assistant deployment with a measured, controlled conversion result, and a range borrowed from somebody else's marketing page is not evidence.

What we do measure is the layer above the widget: which brands and which products AI assistants name when a shopper asks what to buy. Our tracker ran 250 prompts across ChatGPT, Gemini, Perplexity, Copilot and Google AI Mode in the US, UK and India in September 2026 and captured 26,629 citations. Across the 17 non-branded buyer questions in the shopping-assistant category, those answers drew on 1,825 citations spread over 419 domains, and the largest single vendor held 3.4% of them.

That is the number we can stand behind, because we collected it. It measures a different thing from an onsite lift, and it matters for a reason the lift numbers cannot address: an assistant on your site only converts traffic that already arrived. Whether you appear in the answer that sent it is decided somewhere else. That is what Citation Rank measures, and what the House of Zelena case study tracks over six months.

If you want the same reading for your own catalog, the free Citation Rank scan returns it without a call.

Sources

• Rep AI homepage, hellorep.ai, captured 23 September 2026.
• Envive homepage, envive.ai, captured 23 September 2026.
• Alhena AI homepage, alhena.ai, captured 23 September 2026.
• Gorgias, "AI shopping assistants are now officially influencing sales", gorgias.com/research, captured 23 September 2026.
• Webscale AI, "AI Shopping Assistant for Shopify: The 2026 Guide", ai.webscale.com, captured 23 September 2026.
• Retail Systems, "Do AI shopping assistants really lift conversion - or are early gains misleading?", captured 23 September 2026.
• Tru Commerce citation corpus, 26,629 citations, 250 prompts, 5 engines, US/UK/India, captured 16 September 2026.