Most product pages are built for a person who skims, scrolls and clicks. The facts that decide whether the product fits a question can sit in a tab that loads on click, in a lifestyle image, in a PDF, or nowhere. This audit starts from the questions assistants are asked and works back to what a page has to say.
One limit first. Our numbers come from API answers, not the ChatGPT or Perplexity apps. Our data records which pages appeared in source lists, not which sentences an assistant read, so nothing here is a measured effect of on-page copy. The checks come from the questions we ask, from documentation, and from reasoning about how an engine that searches the web and reads the results has to work.
What does each kind of question ask of a page?
In the October 2026 edition, the Dopbot Index asks five prompts per category, one for each intent, and lists them on every category page. Two examples from the checked luggage page: "What is the best checked luggage?" and "best large suitcase under $200". Each type needs different facts to be on the page.
| Question type | What the answer needs from a page | Where pages can fall short |
|---|---|---|
| Best overall | What the product is, who it suits, what stands out | Marketing adjectives with no checkable fact |
| Budget with a price ceiling | Current price and currency, what is included | Price only in an image or set by script |
| Premium | What the extra cost buys: materials, warranty, service | The reason to pay more is implied, not stated |
| Specific use | Dimensions, capacity, compatibility, limits for that use | Specs missing or behind a download |
| Alternatives or buying advice | Criteria and honest trade-offs | No comparison, or a one-sided one |
Why does the page matter more when lists are short? Excluding the alternative-to-a-brand question, which names a brand and inflates counts, budget answers recommended 2.2 brands on average and 93.4% of them recommended one to three (318 answers, October 2026 edition).
Separately, across all answers in that edition, OpenAI's API with web search (gpt-6-luna) linked a mean of 2.0 distinct domains in each of 800 answers, while Perplexity's Agent API with the Sonar model (perplexity/sonar) retrieved a mean of 15.2 in each of 787. These are per-answer averages, not the odds for your page. Brand counts and linked domains are different measures, but both show how few pages can be in play when an engine links only two.
The audit: 12 checks
Score each product page pass or fail. As a rule of thumb of ours, not a measured threshold, start with pages that fail three or more of the first six.
| # | Check | Pass looks like |
|---|---|---|
| 1 | Name and variant | Brand, product, model and variant in the title and first paragraph, spelled as retailers spell them |
| 2 | Price and currency | A current price as text for each variant, with the currency stated |
| 3 | Availability | In stock, backordered or discontinued stated in text, with ship-to regions |
| 4 | Who it is for | One sentence per use you serve, and one on who it does not suit |
| 5 | Specifications | Dimensions, weight, materials, capacity and compatibility as text or a real table |
| 6 | What is included | Contents of the box, accessories and what costs extra |
| 7 | Comparison | How it differs from your other products, and from typical alternatives, stated fairly |
| 8 | Proof | Ratings and review text visible on the page, with the count and the date |
| 9 | Policies | Shipping, returns and warranty terms with specifics, on the page or one click away |
| 10 | Not locked away | No key fact only in an image, a PDF, a click-to-load tab or a script |
| 11 | Parity | Markup and feeds say what the visible page says, including price and availability |
| 12 | Freshness | A visible date on facts that change, and the same facts on every site that lists you |
How do you test a page?
Five tests, from cheapest to most work.
- The source test. Open the page's source (view-source in a browser, or fetch it with a command-line tool) and search for the price and one spec. If they are not in the first response, check 10 fails.
- The no-script test. Turn off JavaScript in the browser and reload. What remains is roughly what a fetcher that does not run scripts sees. Google documents that it renders pages with a headless Chromium once its resources allow. We found no statement on the OpenAI or Perplexity crawler pages about whether their crawlers run scripts, so assume they might not.
- The paste test. Select all the text on the page, paste it into a blank document and ask a stranger to answer eight questions from it: what is it, how much, is it available, what size, what is in the box, who is it for, how do I return it, what does the warranty cover. Unanswered questions are failed checks.
- The match test. Compare name, price and availability across your page, your feed, your Google listing and the retailers that sell you. Any mismatch between your page and your feed is a check 11 failure; any mismatch with your Google listing or a retailer is a check 12 failure.
- The question test. Ask the assistants your baseline questions, three times each, and see whether facts about your product are repeated correctly. Keep the denominator: "2 of 12 answers" is what you have, and it is small.
A minimum fact set from OpenAI's feed spec
As of October 2026, OpenAI's product feed spec lists nine required fields for each item in a feed, which makes a useful floor for a product page. For each purchasable item or variant it requires a stable ID, a title that includes the selected variant when relevant, a factual description, a product page URL with the variant selected where possible, a regular price, an availability value, a main image showing the variant, a brand as shown on the product page and a seller name. It also says to submit one row per purchasable item or variant, and to update prices and availability in the feed when they change.
You can read that list as a page requirement, too. Seven of the nine (title, description, URL, brand, seller name, price and availability) should be stated on the page itself, in text. The ID is an internal key your page does not need to show, and the main image needs alt text and a nearby sentence naming the variant. The spec describes feeds, not crawling, and we have not tested feeds, but a page that cannot supply those facts will struggle to supply a feed or an answer with them.
What belongs in the description, not just the specs table?
Three kinds of sentences carry most of the weight, and they are the ones templates tend to drop.
- A fit statement. "Fits most overhead bins" is weak. "Outside dimensions are X by Y by Z, which fits the carry-on limits listed by these airlines" is checkable. Use your own measurements.
- A reason to choose it. What does the buyer get that they would not from a cheaper or pricier option you also sell? A table of your own range is enough.
- A boundary. Who should buy something else. It makes the other claims more believable, and it supplies the "not for" half of a use-case question.
On Shopify, alt text is described as part of a product description, with 125 characters or fewer recommended. Alt text helps, but it does not replace a sentence in the body. Google says the text near an image can help it understand what the image is about, so the surrounding words are doing the work either way.
What does the own-site data say?
In the October 2026 edition, the linked pages of 46.6% of OpenAI's 800 answers (all five prompts) and the retrieved sources of 62.0% of Perplexity's 787 answers held a domain whose first label matches a recommended brand's name (samsung.com for Samsung). The match misses country domains, parent companies and other store names, and Perplexity's lists are far longer than OpenAI's, so do not compare the two figures. That shows brand sites appear next to answers. It does not show that any page, or any sentence on it, drove a recommendation, and a retrieved Perplexity page may not have been used at all. Treat the audit as removing reasons a page could fail to help, not as a ranking factor list.
What this audit does not cover
- Markup. The structured data guide covers Product, Offer and the rest.
- The apps. Our data is API answers, and the apps can show different pages and use different retrieval.
- Other pages. Category, comparison and policy pages matter too; this audit stays on product pages.
For a store-wide sequence, the 30-day Shopify plan puts this audit in context. The methodology explains how we count.
Frequently asked questions
Should I put the same text on every variant page?
Write the facts that are true of the whole product once, and make each variant's price, availability and options specific to it, whether variants share one page or have their own. OpenAI's feed spec asks for one feed row per variant, with the selected variant in the title and URL where relevant or possible.
Is a long description better than a short one?
We have no data on length. As of October 2026, Shopify's help page for its ChatGPT channel says product descriptions should include any relevant legal disclosures in the first 6,000 characters. It does not say what happens to text after that, so put the facts a shopper needs early as a precaution. Beyond that, include every fact a shopper's question could need and stop.
How we write: drafted with AI tools from Dopbot Index data and public sources; every number and outside fact is traced to its source in a claims record and checked before publishing. More about how we write; corrections to hello@dopbot.com.