How to get your products recommended by ChatGPT: a practical checklist.

There is no form to fill in and no fee to pay to be recommended by ChatGPT. There is, though, a fairly predictable list of things that make a product easy for it to find, check and stand behind. This is that list, in the order we work through it with clients.

We have written it for e-commerce teams: a head of e-commerce, a developer and whoever owns content and PR. ChatGPT is the main subject, but the same steps apply to Gemini, Perplexity, Claude and Meta's Muse, and a section below covers where they differ.

How does ChatGPT decide which products to recommend?

ChatGPT recommends the products it can find, match to the question and verify, drawing on product feeds, web pages and independent sources. OpenAI does not publish a ranking formula, and you should be wary of anyone who claims to have one. What can be observed from the outside is consistent, though: answers favor products whose price, availability and key attributes are stated clearly and agree across sources, and that reviewers or buying guides have written about.

Recommendations are not adverts. You cannot buy a place in the organic answer. The shortlist an agent builds, which we call the candidate set, is earned through legibility and reputation. That is good news for brands with a strong product and tidy data, whatever their size. Google, though, is now scaling ads inside AI Mode, and we expect fewer free slots over time; see what Google's and Reddit's earnings mean for stores.

What does our own data show?

Our data shows a short shortlist, and agents that disagree about who is on it. For edition one of the Recommended by Agents Index, we asked ChatGPT, Gemini, Claude, Meta's Muse and Perplexity the same forty shopping questions about running shoes under $150 in the US. That gave 200 answers, collected on September 28, 2026. We counted every brand each agent put forward as something to buy.

• The shortlist is short. Each answer recommended about four brands. Seven brands took 83% of all recommendations, and twenty other brands shared the rest.
• ChatGPT has its own favorites. ChatGPT recommended Saucony in 75% of its 40 answers and ASICS in 73%, ahead of Brooks at 50%. Across all five agents, Brooks led with 67%.
• Agents disagree. HOKA was recommended in 65% of Gemini's answers and 13% of ChatGPT's. Puma was recommended in 35% of ChatGPT's answers and in none of Perplexity's.
• The budget in the question is a filter. On was recommended in 6 of 200 answers. Much of its range sits above $150, so the agents mostly left it out.

Three things follow for the checklist. First, there is no page two. A brand outside the four or so names in an answer is not ranked lower; the shopper never sees it, which is why the data and citation work in steps 4 to 11 matters. Second, a brand's standing in one agent says little about another, so measure each one separately (step 12). Third, a budget in the question works as a filter, so a price the agent can confirm (steps 4 to 6) is a condition of entry, not a detail.

What this data cannot tell you. It covers one category, one market and one answer per question per agent, so treat gaps of a few points between brands as noise. The answers came through each provider's developer API with web search on, not the consumer apps, which can differ in default model, search stack and personalization. And it shows which brands were recommended, not why: nothing here proves that any single fix moves a brand into an answer. Our full methodology asks each question five times per agent; this first edition is smaller than that.

What is the checklist?

The checklist has twelve steps in four stages: find out where you stand, make the product machine-readable, earn independent citations, and measure. Do them in order. Each stage makes the next one worth doing.

Stage 1: Find out where you stand

1. Ask the questions your customers ask. Write ten to twenty full-sentence questions for your main category, each with a budget, a use and a constraint. Ask each one several times in a clean ChatGPT session and note which brands are recommended. Answers vary from run to run, so one pass is not a baseline. Our running shoes walkthrough shows the method, and the forty questions on the Agent Index page are a worked example.
2. Check what it says about you directly. Ask "What does [your brand] sell?", "Is [your product] good for [use]?" and "What is [your brand]'s return policy?" Wrong or vague answers point straight at missing or conflicting data.
3. Check crawler access. Look at your robots.txt and your CDN or firewall rules. OpenAI documents separate user agents for search indexing (OAI-SearchBot), for fetching pages during a conversation (ChatGPT-User) and for model training (GPTBot). Blocking the first two can keep you out of answers. Whether to allow training crawlers is a separate business decision.

Stage 2: Make the product machine-readable

4. Publish complete Product schema on every product page. At minimum: name, description, image, brand, SKU and GTIN where you have one, plus an Offer with price, currency, availability and URL. Add AggregateRating and Review if you have genuine reviews. Print it in the HTML the server sends, validate it, and make sure it matches what the page visibly says. In our audit of twelve DTC stores, ratings and GTINs were the most common gaps, missing from 8 of 11 scored pages each.
5. Add shipping and returns as data. Use shipping details and a merchant return policy in your markup, and state both in plain text on the product page. "Free returns within 30 days" is the kind of sentence an agent can repeat to a shopper. A link to a PDF is not.
6. Keep a product feed in step with your store. ChatGPT's shopping features can take merchant product feeds under OpenAI's Agentic Commerce Protocol (ACP), and Google's surfaces use Merchant Center. Whatever feeds you run, the rule is the same: the price and stock in the feed must match the live page. A stale feed is worse than none, because it creates a conflict.
7. Write product copy that answers questions. Put the facts shoppers filter on into text: who it is for, what it is made of, dimensions, weight, compatibility, care. Use the words shoppers use. "Fits wide feet" beats "generous last".
8. Make the page readable without a fight. Price, stock and variants should be in the first render, not injected late by a script. Reduce pop-ups to the consent banner you are required to show. This matters most for browser agents, as we explain in why Muse reads your site like a customer.
9. Add an llms.txt file, with modest expectations. It is a simple, low-cost summary of who you are and where your key pages live. It is a proposed convention, not a ranking factor anyone has confirmed. Details in what AI shopping agents actually read.

Stage 3: Earn independent citations

10. Map the sources ChatGPT quotes in your category. When it answers your baseline questions with web search on, look at the sources it links. You will usually see a small, repeating group of buying guides, review sites, publications and forum threads. That is your target list.
11. Earn your place in those sources. Send products for review, supply accurate specifications to guide editors, answer questions on Reddit and forums openly as the brand, and build a base of detailed, recent customer reviews. Do not fake any of it. Agents cross-check, and so do the communities involved.

Stage 4: Measure

12. Re-run your questions on a schedule. Monthly at least, in ChatGPT and at least one other agent. Record whether you were recommended or only mentioned, and where, next to two or three named competitors. Report it as share of recommendation: the percentage of answers that put you forward as something to buy. Changes take weeks to show, so keep the question set stable and watch the trend.

Who should own each step, and how long does it take?

Ownership splits across three roles, and most of the technical work is a one-off project while citations and tracking are ongoing. The table below is a planning guide, not a promise; effort depends on your platform and the number of page templates.

What are the most common mistakes?

The most common mistake is treating this as a content trick instead of a data-quality project. The others follow from it.

• Writing for the model instead of the shopper. Pages stuffed with "best [product] recommended by AI" read badly to people and do nothing for agents, which are matching facts, not slogans.
• Letting sources disagree. A sale price on the page, the old price in the feed and a third price in the schema. When sources conflict, the agent's safe option is to recommend someone else.
• Blocking agents by accident. Bot-protection defaults that challenge every automated visitor, including the ones sent by shoppers.
• Judging by one answer, or one agent. Answers vary from run to run, and agents disagree: in our Index, Puma was recommended in 35% of ChatGPT's answers and never by Perplexity. One screenshot, good or bad, is an anecdote.
• Skipping the open web. Perfect schema with no independent coverage gives an agent facts but no reason to prefer you.
• Expecting overnight results. Feed and schema fixes can show up quickly. Citations take longer; we tell clients to expect six to ten weeks before earned coverage shows in what agents say.

How does this apply to Gemini, Perplexity, Claude and Muse?

The same twelve steps apply to every AI shopping agent, because each one has to find, match and verify a product. What changes is which version of your product it leans on: the feed, the open web or the live page. That decides which stage repays the most work. ChatGPT's shopping features take merchant feeds, and its answers also draw on web search, so it needs all three stages.

Gemini is feed-led, like ChatGPT's shopping features, so Stage 2 matters most: a complete, current Merchant Center feed that matches the page. One Index note: through the API, the newer Gemini 3.8 Flash answered from memory without searching the web, so we used 3.5 Flash, which searched on every question. When a model answers from memory, only what was already written about you can count.

Perplexity and Claude lean on search results and on what reviewers, guides and forums say, so Stage 3 carries more weight. Step 3 matters too: Anthropic and Perplexity document their own crawlers and user-triggered fetchers, as OpenAI does, and a blanket block shuts them out. In our Index, Perplexity was the most selective agent, recommending fewer than three brands per answer on average. Claude spread its picks widest, recommending 22 different brands at least once, against 17 for ChatGPT. With one answer per question, treat that as a pattern to watch, not a rule.

Meta's Muse is a browser agent. It can open your live store, read the page through the accessibility tree, pick a size and try the checkout, so step 8 matters most, followed by delivery and returns in plain text (step 5). Shopify merchants also reach Muse through Shopify Catalog. In our Index, Muse recommended Brooks in 37 of its 40 answers. We asked Muse's model through its developer API, so the Index measures what it recommends, not whether it can use your store. The ten-minute browser-agent test covers that.

That is why we measure across five agents in the Agent Index rather than one. A brand can lead in one agent and be missing from another, and the gap tells you which stage to work on.

Where should you start this week?

Start with steps 1 to 3, which cost nothing but a day. You will know whether you are in the candidate set, what ChatGPT gets wrong about you, and whether you are blocking it. If you want a second opinion, the free check scores one product URL and shows the top three gaps. The full Audit covers every template, every feed and all five agents, and the price is on the page. Or book a thirty-minute call and we will tell you where we would start.

Written by Harry Harrison, founder of Recommended by Agents, under our editorial standards. Found an error? Tell us. Published September 15, 2026; updated October 31, 2026 with results from the Agent Index and a section on other agents, and November 4, 2026 with a question on Gemini. Spotted something out of date? Tell us and we will fix it.