Why doesn't ChatGPT recommend your products? A playbook for e-commerce brands

Is AI shopping traffic big enough to matter yet?

Big enough to plan for, too small to panic about. AI answers send a small share of shopping visits today, but that share is growing quickly and the visitors who arrive buy more often than average. The brands an engine names on your category prompts are the shortlist those shoppers see first.

Fast growth, better conversion

Adobe Digital Insights measures traffic to US retail sites, and its April 2026 report has the clearest retail numbers we know of. AI-referred traffic to US retail sites grew 393% year on year in Q1 2026. In March 2026 those visitors converted 42% better than other traffic and earned 37% more revenue per visit.¹

Two cautions before these go into a board deck. They are US figures, and they count visitors who clicked through from an AI answer, not every shopper who asked one. Many shoppers read the shortlist, close the chat and search for a brand by name later. Our view: the conversion gap is the more useful number, because it says the answer has done some of the selling before the visit starts.

US retail, Adobe

AI-referred shoppers arrive ready to buy

+393%AI-referred visits to US retail sites, year on yearQ1 2026¹

+42%conversion compared with other trafficMarch 2026¹

+37%revenue per visit compared with other trafficMarch 2026¹

Still a small share of visits

The growth starts from a low base. SE Ranking, which sells SEO and AI visibility software, tracked 101,574 websites in 250 countries and territories. AI search sent 0.32% of all their visits in January–April 2026, against 0.02% in 2024.² In the same study, ChatGPT sent 74.78% of AI referral traffic in January–April 2026, down from 79.74% in 2025, and Gemini sent 11.56%.³

Similarweb counts an average of 770.7 million AI referral visits a month worldwide between June 2025 and May 2026.⁴ ChatGPT is still where most AI referrals start, but its lead is narrowing. A plan built for ChatGPT alone will age fast, so run the same prompts in Google AI Mode, Gemini and Perplexity as well.

Where does an AI shopping answer come from?

From two places: product data the engine can parse, and what other people have written about the product. Your own pages sit between the two, and they help most when they agree with both.

What OpenAI says it reads

OpenAI's help page for shopping in ChatGPT says product results consider "structured metadata from first-party and third-party providers (e.g., price, product description) and other third-party content".⁵ The same page says Shopify merchants' product data already reaches ChatGPT through Shopify Catalog, and that other merchants can apply to send OpenAI a direct product feed.⁵

That gives a brand two levers. The first is data: price, description, stock and variants, in a form a machine can read without guessing. The second is everything written about your product away from your site. The page does not say how the two are weighed. Be wary of anyone who claims to know the formula.

What the citation data adds

Profound, which sells AI visibility tracking, analysed 11.84 billion citations from eight engines between 16 April and 16 July 2026. The data covers all 8,061 categories it tracks, not shopping alone. About 57% went to company-owned websites: 47% on ChatGPT and 69% on Gemini.⁶ So your own pages matter, and on ChatGPT more than half of citations still point somewhere else.

Our view: for shopping prompts, somewhere else usually means review sites, editorial buying guides, marketplace listings and community threads. That is our judgement, not a published finding. An audit tests it category by category instead of assuming it.

How a shopping answer is assembled

1. Shopper prompt
2. Engine searches
3. Product data and feeds
4. Brand pages
5. Reviews, guides, marketplaces
6. Shortlist of brands

Cause 1: the engine cannot read your product data

If price, stock and variants are missing, stale or locked inside images and scripts, the engine has less to match against a shopper's request than it has for a competitor with a complete feed. This is the cheapest cause to find and the cheapest to fix.

The fields that do the work

The check takes an hour. Export your feed and read ten rows the way a machine would. Is there a price, availability, size, colour or width, a brand or GTIN, and a factual description on every row? Then open a product page with scripts turned off and see what text is left.

Where schema fits

Schema is hygiene, not the lever. In an Ahrefs test of pages that were already heavily cited, run from August 2025 to March 2026, adding schema produced no measurable uplift in citations on ChatGPT or Google AI Mode.⁷ The test covered pages engines already cited, so it says less about pages they ignore. We set out that evidence in our post on schema, llms.txt and Reddit.

Our view: mark up Product and Offer correctly, because it is cheap and it keeps your facts in one consistent shape. Then put the real effort into the feed. A feed with every variant and live stock does more than perfect markup on a page that shows last month's price.

Cause 2: do the sources it trusts mention you?

Often they don't, and this is the cause most brands miss. If the review sites, buying guides and threads an engine reads for your category list three competitors and not you, a clean feed will not get you onto the shortlist. You have to be in the places the answer is built from.

Find the pages that name your competitors

Run your category prompts in engines that show their sources, such as Perplexity, ChatGPT with search and Google AI Mode. Copy every cited link into a sheet and tag it: brand site, marketplace, review site, editorial guide, forum or video. Then mark which ones name your competitors and which name you.

That sheet is your target list. It is also the most useful thing a first audit produces, because it turns "be more visible" into named pages.

Earn a place on them

Each source type needs a different approach. Editors who maintain buying guides update them, so send product samples with the facts they compare on. Make marketplace listings as complete as your own site. Ask real customers for reviews at the moment they are happiest, through the channels you already run.

In community threads, answer as the brand and say so. Do not write reviews yourself, seed threads or buy placements dressed up as editorial. Engines and shoppers both read those pages, and a placement that is found out costs more trust than it bought. This is the slow part of AI visibility work, and it takes months rather than days.

Cause 3: your product facts disagree across the web

When your site, your feed, a marketplace listing and an old review give different prices, sizes or ingredients, the engine has to choose one. It may choose the wrong one, or hedge and recommend a competitor whose facts agree everywhere. If an engine already repeats a wrong fact, our guide to correcting wrong AI answers covers the trace, the report and the recheck.

The check: list ten facts for each of your best-selling products, such as price, sizes, materials or ingredients, warranty, returns window and the countries you ship to. Compare them across your site, your feed, your top marketplaces and the three reviews cited most often. Correct everything you control the same week. For third-party pages, ask the publisher for an update and give them the dated change.

The four causes on one page

One product can have more than one cause at once, so check all four before you spend on any of them. The table below is the short form of this playbook.

Illustrative: the checklist we work from in an audit. Owners vary by team size.

Cause 4: is checkout even live in your market?

Outside the United States, mostly not, as far as the engines publish. Checkout inside AI answers is documented mainly for US shoppers, and most engines say nothing about other markets. We read each engine's own help or launch page on 6 October 2026 for the markets this playbook covers.

Checkout inside AI answers, as published

Checked 6 October 2026 on each engine's own page. "Not published" means the page does not say; it does not mean the feature is absent. Sources 8 to 13.

Google's page is inconsistent. One section says agentic checkout works in the US and Australia; another says it is available only in the US.⁹ Google's Merchant Center help lists UCP-powered checkout eligibility for the United States, Canada and Australia.¹⁰ On 20 May 2026, Google said that checkout would roll out across Canada and Australia "in the coming months".¹¹

Product answers reach further than checkout

Being recommended does not depend on checkout. OpenAI's shopping help page gives no country list for product results.⁵ In October 2025, Google said AI Mode would be available in over 200 countries and territories.¹⁴ That announcement is now a year old, so check Google's current list for your market before you rely on it.

So Causes 1 to 3 apply wherever your shoppers ask; checkout only changes where the sale closes. Our view: outside the US, plan for answers that send shoppers to your site or marketplace listing, and treat in-answer checkout as a US experiment until the engine's own page says otherwise.

How to audit your AI shelf in a week

Pick the prompts your shoppers would type, run them in six engines more than once, and record every brand and every cited source. A week is enough for a first picture of one category, if one person owns it.

Build the prompt set

Mix four kinds of prompt: category ("best X for Y"), problem ("what helps with Z"), comparison ("brand A or brand B") and brand ("is brand A worth it"). Write them the way shoppers talk, and write them for each market you sell in, because answers differ by country.

Run it more than once

The same prompt can return a different shortlist on a second run, so one screenshot proves little. Run each prompt several times per engine and count the share of runs that name you. Our measurement spec for AI visibility scores explains how many runs that takes. Our paid audit does this at scale: 150 buyer prompts across six engines, each on three separate days, with confidence intervals against five competitors.

An AI-shelf audit, one day at a time

1. Day 1: build the prompts

   Category, problem, comparison and brand prompts for each market.

2. Day 2: run them

   Several runs per prompt in each engine, saved with the date.

3. Day 3: record brands and sources

   Who is named, and every cited link, tagged by type.

4. Day 4: check your own data

   Feed rows, scripts-off pages and the ten-fact comparison.

5. Day 5: rank the fixes

   Owner and order for each cause you found.

What should you fix first?

Fix the feed and your product facts first, because you control them outright and nobody has to agree to the change. Then start the slower work of earning mentions, since it compounds over months.

Fixes you control

Complete every feed field for your best sellers, with live price and stock. Make sure price, sizes and key specs are plain text on the page. Bring marketplace listings into line with your site. Add a short, factual line to each product page on who the product is for and who it is not for, since shopping prompts are usually phrased as a need.

Most of these changes sit with one team and need no outside approval. Do them for the twenty products that sell most before touching the long tail, then re-run your prompt set so you have a before and after on the same prompts.

Work you have to earn

Reviews, buying guides and marketplace reputation take longer. Start a review-request programme with real customers, pitch the guides on your target list, and answer honestly in the threads where shoppers ask. If you want us to run the diagnosis, the AI Visibility Audit is $3,500 (₹1.5L) at our published prices.

Where to start on the AI shelf

Effort →

Big betsEditorial guides and a review programme.

SkipPaid listicle placements and fake reviews.

Quick winsComplete feed fields and consistent product facts.

Fill-insSchema tidy-up and special AI files.

Impact →

When a where-is-my-order agent pays for itself

When order-status and returns questions are a large, repetitive share of your support tickets and your order data is clean. More shoppers from AI answers means more of those questions, and a support agent grounded in your Shopify or WooCommerce order data can answer them on web, WhatsApp and email, then hand the rest to your team.

Our view: build it after the shelf work, not before, and only if the ticket mix supports it. It is scoped as an automation sprint that starts at $10,000 (₹4L), our published price, with deflection and first-response time measured and signed off.

The short version

ChatGPT leaves a product out for one of four reasons: it cannot read the product data, the sources it trusts do not mention the brand, the facts disagree across the web, or the feature is not live in that market. Each has a check that takes hours, not weeks.

Start with what you own, the feed and the facts. Then build the list of pages that name your competitors and earn a place on them. Measure with repeated runs in every market you sell in, and expect the third-party work to take months. No engine publishes how long a change takes to show.