To find out whether ChatGPT can find your products, ask it five things: what your brand sells, what one exact product costs, what it reads when it opens your product page, which brands it recommends for a real shopper's question, and which sources it quotes. The whole test takes about ten minutes, and every failure maps to a fix.
This is the quick version of the baseline we run at the start of every Audit. It won't give you a score you can put in a board deck, but it will tell you whether you have a problem, and roughly where. If you want the full method afterwards, our guide to getting your products recommended by ChatGPT covers all twelve steps.
Before you start
You need ChatGPT, one product page from your store, and a note of the true facts to check against: price, stock, delivery cost and return window. Set up so you see what a stranger sees, not what ChatGPT has learned about you.
- Use a temporary chat, or sign out. A temporary chat doesn't use or update ChatGPT's memory of you, so your past conversations don't steer the answer.
- Make sure it searches the web. If an answer has no sources, ask again with “search the web” at the start. Without search, ChatGPT answers from what it learned in training, which can be months old.
- Pick a product that matters: a best seller with a clear price, not a gift card or a bundle.
- Write down every answer. A screenshot per test is enough. You will want to compare next month.
Test 1: Does ChatGPT know who you are?
Ask: “What does [your brand] sell, and what is its return policy?” Time: two minutes.
A pass is a correct summary of what you sell and your actual return window, with your own site among the sources. A fail comes in three kinds, and each means something different:
- It confuses you with another company. Your brand name is ambiguous, and nothing on the web ties it firmly to your store. Organization schema on your home page, with your logo and social profiles, helps it pick the right one.
- The return policy is wrong or vague. The policy is probably not on your site as plain text, or it disagrees between pages. A PDF or an image of the policy doesn't count.
- It cites resellers or old press instead of you. Your own pages aren't being found or trusted. Start with Test 3.
Test 2: Can it find one exact product?
Ask: “How much is the [exact product name] from [your brand], and is it in stock in [a size or color]?” Time: two minutes.
A pass is today's price, the right stock status and a link to your product page. ChatGPT tags its links with utm_source=chatgpt.com, so a click from this answer will show up in your analytics; our GA4 guide shows where.
- An old price usually means two of your sources disagree: the page, the Product schema in its code and your product feed. When they conflict, an agent can't tell which is right. Generate all three from the same data.
- A link to a marketplace or reseller means the agent found a version of your product it trusts more than yours, often because it has clearer data or reviews.
- “I couldn't find that product” means the page is hard to find or hard to match. Check the product name on the page is the name shoppers use, and that the page is in your sitemap.
Test 3: Can it read your product page?
Ask: “Open [your product URL] and tell me the price, the delivery cost and the return window.” Time: two minutes.
When you give ChatGPT a link, it fetches the page with a user agent called ChatGPT-User, which OpenAI documents alongside OAI-SearchBot for search and GPTBot for training. This test shows what that fetch gets back.
- “I can't access that page” almost always means something is blocking it: a rule in robots.txt, or more often a bot-protection setting at your CDN or firewall that challenges every automated visitor. Fix it in the bot settings, not by turning protection off.
- It opens the page but misses the price, or gets the variant wrong. The price is probably drawn by a script after the page loads. Put price, stock, delivery and returns in the HTML your server sends.
- It gets delivery or returns wrong. Those facts live behind a tab, a pop-up or a separate page. State them on the product page in a sentence.
Browser agents such as Meta's Muse go further and try to use the page like a shopper. If you sell to them, run our browser-agent test too.
Test 4: Are you on the shortlist?
Ask one question the way a shopper would, with a budget, a use and a constraint, in three fresh temporary chats. For example: “What are the best trail running shoes under $130 for wide feet?” Time: three minutes.
Count how many of the three answers recommend you, as something to buy, not just mention you. Note who else is recommended every time.
Why three times: answers vary from one run to the next. One good answer is luck; two out of three is a pattern. And the shortlist is short. In edition one of our Agent Index, AI agents recommended about four brands per answer across 200 answers, and seven brands took 83% of all recommendations. If you aren't in the four, the shopper doesn't see you. We call that shortlist the candidate set.
- Recommended in two or three of three: you are in the candidate set for that question. Protect it and track it monthly.
- Mentioned but not recommended: the agent knows you but has less reason to back you than the others. Usually that means fewer reviews or less independent coverage.
- Absent: if Tests 1 to 3 passed, the gap is almost always off your site. Test 5 shows where.
Test 5: Which sources does ChatGPT trust?
Do: on one of the Test 4 answers, open the list of sources. Time: one minute.
You will usually see a small group of buying guides, review sites, publications and forum threads. That is the list ChatGPT is drawing its shortlist from for your category, and the brands it recommends are the ones those sources write about. Note whether you appear in any of them. If you don't, that list is your outreach target: send products for review, give guide editors accurate specifications, and answer questions openly as the brand where shoppers ask them.
The scorecard
Use the table to turn your five results into a to-do list. Work top to bottom: a page ChatGPT can't open makes every other fix pointless.
| Test | Pass looks like | A fail usually means | Start with |
|---|---|---|---|
| 3. Open the page | Reads price, delivery and returns correctly | Bot rules block it, or facts load by script | Bot settings; facts in the server-sent HTML |
| 2. Exact product | Today's price, your page linked | Page, schema and feed disagree | One data source for all three |
| 1. Brand facts | Correct summary and policy, your site cited | Policies not in plain text; ambiguous brand | Policy text; Organization schema |
| 4. Shortlist | Recommended in 2 of 3 answers | Too little independent evidence | Reviews and coverage |
| 5. Sources | You or your reviewers are cited | You're absent from the sources it trusts | Outreach to those sources |
Does this work for Gemini and Perplexity?
Yes. Run the same five checks in each, because agents disagree. In our Index, HOKA was recommended in 65% of Gemini's answers and 13% of ChatGPT's, and Puma in 35% of ChatGPT's and none of Perplexity's. A brand can pass in one agent and be missing from another.
The emphasis shifts a little. Gemini leans on Google Merchant Center product data, so Test 2 matters most there. Perplexity leans on search results and reviewers, so Tests 4 and 5 tell you more. For a Google-side view without asking anything, Search Console now has a generative AI report that shows your impressions in AI Overviews and AI Mode.
What the test can't tell you
It is a smoke test, not a measurement. Three runs of one question is a tiny sample, and the ChatGPT you use in a browser can differ from what another shopper sees on a different plan, in another country, or with memory on. It also shows what ChatGPT says, not why: nothing here proves that a single fix will move you onto the shortlist.
For a number you can track, keep a fixed set of ten to twenty questions, ask each several times a month in at least two agents, and report the share of answers that recommend you. That is how our Index methodology works, and it is what the Track service does for clients.
If you would rather start with the page itself, the free check scores one product URL against seventeen things agents read and shows your top three gaps.
Key takeaways
- Five checks in ten minutes: brand facts, one exact product, opening your page, a shopper's question three times, and the sources cited.
- Use a temporary chat with web search on, so you see what a stranger sees.
- Fix in order: access first, then consistent product data, then policies in plain text, then independent coverage.
- The shortlist is short: about four brands per answer in our Index. Being mentioned is not the same as being recommended.
- Agents disagree, so repeat the test in Gemini and Perplexity, and re-run it monthly.
Questions and answers
How do I check if ChatGPT recommends my products?
Ask it a question the way a shopper would, with a budget, a use and a constraint, in three fresh temporary chats with web search on. Count how many answers recommend you as something to buy. Two out of three means you are on the shortlist for that question.
Why does ChatGPT show the wrong price for my product?
Usually because your sources disagree. The visible page, the Product schema in its code and your product feed can each carry a different price, especially after a sale starts or ends. Generate all three from the same data so they always match.
Why can't ChatGPT open my website?
Most often a bot-protection setting at your CDN or firewall challenges every automated visitor, including ChatGPT-User, the agent ChatGPT uses when someone shares a link. A rule in robots.txt can also block it. Adjust the bot settings rather than switching protection off.
Should I sign out of ChatGPT before testing?
Sign out or use a temporary chat. A temporary chat doesn't use ChatGPT's memory of your past conversations, so the answer is closer to what a new shopper would see. It is still only one view, so treat it as a quick check.
How often should I run this test?
Once a month, with the same questions, so you can compare. Answers vary from run to run and agents change often, so the trend over several months tells you more than any single result.
Written by Harry Harrison, founder of Recommended by Agents, under our editorial standards. Found an error? Tell us. Published . Index figures are from edition one, collected September 28, 2026. OpenAI, ChatGPT, Gemini, Perplexity, Meta and other names are trademarks of their owners, and we have no affiliation with them.
