Discussion starter: an AI search question, posted with an answer from Niraj Raut to open the thread. If you have dealt with this on a site, reply with what you saw, especially where it differs.
- AI Search
When AI answers misdescribe your product, what actually corrected it: your own pages, third-party reviews, Wikidata, or nothing you could control?
Disclosure: Niraj Raut, who posted this answer, runs the SEO consultancy linked at the end of it.
1 reply
-
adminAdmin
What corrects it is whichever source the answer engine was actually reading when it got your product wrong, and that differs by platform and by prompt. In search-grounded answers (AI Overviews, AI Mode, ChatGPT with search, Perplexity, Copilot), the wrong claim comes either from a page the system retrieved, yours or a third party’s, or from the model itself: blending pages that were each correct, or filling a gap from memory. The first kind is fixable at the source once that page is recrawled. The second needs clearer wording on the pages being read. When the answer comes from the model’s memory, with no retrieval at all, nothing you control changes it quickly.
Wikidata is the narrowest of the levers in the question. It is a reasonable place to fix company-level facts such as founder, headquarters or parent company. It does little for pricing, plans and feature lists.
Public evidence on which lever works is thin. What has been published on correcting AI answers comes mostly from vendors and agencies selling AI visibility tools or services, and none of it that I can point to isolates one lever under controlled conditions. AI answers also vary between runs of the same prompt, so a correction judged from a handful of screenshots is weak evidence either way. What follows is the mechanism, a way to trace each error to the layer that produced it, and a measurement routine that makes a before-and-after worth believing.
Four layers where an answer can go wrong
Each layer has a different owner and a different fix. Trace first and edit second, because an edit aimed at the wrong layer produces no change, and then a false conclusion that nothing works.
Layer How it shows up What you control When a fix can show Retrieval The answer cites or links sources, and the wrong claim appears on one of them The cited page: edit yours, request a correction on a third party’s. Also crawler access and snippet controls After that page is recrawled and retrieved again; sooner for pages crawled often Synthesis Every cited source is correct, but the answer blends them (a competitor’s feature attributed to you, two plans merged, a same-name company conflated) or fills a gap from memory The wording of the pages being retrieved After recrawl, with no guarantee Model memory No search ran (search off, or a mode that answers without browsing), and the claim matches how you were described before the model’s training cutoff Nothing directly. Over time: how consistently the web describes you, and whether training crawlers can read your site With a later model version, not after a recrawl Entity data Company-level facts are wrong (founder, headquarters, founding year, parent company), sometimes matching an error in a knowledge panel Organization markup and the About page on your site, Wikidata if the company qualifies, knowledge panel feedback Slow, and it varies by system What each lever can and cannot fix
Your own pages: the only source you can change today
Start with your own estate, because stale first-party content is a source you can fix this afternoon. Search it for the wrong claim’s wording: old pricing pages, legacy help-centre articles, launch posts for features you have since removed, PDF datasheets, your own press releases (including copies on wire services), app store listings, reseller and partner pages, and Product or Offer markup still carrying an old price. Update or redirect them, and add a dated note where an old post has to stay up. Fix stale markup for its own sake: Google’s AI features documentation lists structured data that matches the visible text among its best practices, while saying there is “no special schema.org structured data that you need to add” for these features.
Then make sure you have not removed yourself as a candidate source. The same documentation says a page must be indexed and eligible to be shown in Google Search with a snippet to appear as a supporting link in AI Overviews or AI Mode, and lists
nosnippet,data-nosnippet,max-snippetandnoindexas the controls that limit what is shown. A pricing table wrapped indata-nosnippetis a pricing table Google has been told not to show. OpenAI’s crawler documentation separates OAI-SearchBot, used for ChatGPT’s search features, from GPTBot, used for training data. A robots.txt that blocks OAI-SearchBot limits what ChatGPT search can read from your current pages. Blocking GPTBot for licensing reasons is a legitimate choice, but future models then have less of your own wording to learn from.The same logic applies to Google’s site-level Search generative AI control. As of September 2026 it is available to all websites, after a worldwide rollout on 31 August 2026, and opted-out sites get no traffic or impressions from AI Overviews, AI Mode or AI Overviews in Discover. Opting out removes your pages as a source. It does not remove the questions, so answers about your product then have to draw on other sites.
Finally, state the fact the answers get wrong in plain words, in the HTML text of the page most likely to be retrieved for that question. If answers say you have a free plan, the pricing page should say “There is no free plan” rather than leaving readers, and models, to infer it from the absence of one. Put it in text rather than in an image or a script-built pricing widget, since not every crawler runs JavaScript. That is ordinary clear writing, not the artificial “chunking” that Google’s guide to generative AI features says you do not need. A visible “as of” date also helps readers, and gives any system comparing conflicting sources a way to tell which one is current.
Third-party reviews and comparison pages
When the cited source is a review platform, a comparison listicle or a forum thread, nothing on your site touches it. The levers are slower and partly outside your control: update the profile fields you manage on review platforms, send publishers a short correction request (one fact, the evidence, the date it changed) rather than a request to rewrite their verdict, and answer forum threads openly as the company. Some publishers never revisit old articles. Then the aim shifts to making current, better sources available for retrieval, including fresh reviews and your own comparison pages. The same Google guide says seeking inauthentic mentions is not helpful, and seeded reviews or planted threads carry reputational and legal risk of their own.
Wikidata
Wikidata is worth fixing when the error is about the company as an entity, the company meets Wikidata’s notability rules, and every statement you add carries a reference. Its data is openly licensed and widely reused, so an error there can spread. It is a weak lever for product descriptions, because pricing, plans and feature lists are not what Wikidata items are built around, and whether a given assistant consults Wikidata at answer time is not something you can observe. No public evidence that I can point to shows a Wikidata edit, on its own, correcting a chatbot’s description of a product.
Product feeds, if you sell physical goods
Sellers of physical products have one more first-party channel. Merchant Center data feeds Google’s shopping experiences, and Google’s shopping updates from Google Marketing Live added conversational attributes to Merchant Center, plus AI performance insights with share of voice, which as of September 2026 are available for English-language queries in Australia, Canada, India, New Zealand and the US. If AI shopping answers get a product’s specifications wrong, check the feed before the product page.
What sits outside your control
Model memory, until the platform ships a model trained on newer data. Synthesis quirks that survive clear wording. Third-party pages that will not change. And run-to-run variation, which means a corrected answer can still come back wrong in some runs. For all four, the remaining move is to make the correct version easier to retrieve than the wrong one, then measure over time rather than per screenshot.
Trace the error before you edit anything
- Pin down the claim. Write the wrong statement and the correct fact, with the date the fact changed. A model trained before that date cannot have learned the new version.
- Sample, do not spot-check. Build a fixed set of prompts in customers’ words: pricing, limits, integrations, comparisons, “is X good for Y”. Run each several times per platform in clean sessions (logged out or memory off, same location), and log the answer, whether a search ran, and every cited URL. SparkToro and Gumshoe’s January 2026 research (600 volunteers running 12 prompts, 2,961 runs across ChatGPT, Claude and Google’s AI) concluded “there’s a <1 in 100 chance that ChatGPT or Google’s AI, if asked 100X, will give you the same list of brands in any two responses”. That study measured brand lists, not product descriptions, but the lesson carries over: record an error rate per platform, not a screenshot.
- Classify each wrong answer by layer. A cited page contains the claim: retrieval. Cited pages are correct but the answer is not: synthesis, which in practice you treat the same way whether the model blended sources or filled a gap. No search ran: memory. Company facts wrong: entity data.
- Fix at source, your own domain first. Then request recrawls: URL Inspection for Google, and IndexNow or URL submission in Bing Webmaster Tools for Bing, which Copilot draws on.
- Confirm the corrected pages are being surfaced. Search Console’s Search generative AI performance report, available to all eligible sites as of September 2026, shows impressions by page for AI Overviews, AI Mode and generative AI features in Discover, with no clicks or queries. Bing Webmaster Tools’ AI Performance report, launched as a public preview in February 2026, shows page-level citations and a sample of grounding queries across Copilot and Bing’s AI summaries. Neither shows what the answer said, so pair them with your prompt log.
- Re-run the same prompt set once the changed pages show a crawl date after your edit (URL Inspection shows Google’s last crawl), then again a few weeks later, because retrieval sets and model versions shift underneath you. Compare error rates per platform and per layer.
Hypothetical example:
A SaaS company dropped its free plan in March. In September, 20 runs per platform show “has a free plan” in most ChatGPT answers where no search ran, in several Perplexity answers citing a 2025 comparison listicle, and in some AI Mode answers citing the company’s own 2024 launch post. The fix sequence: add a dated note to the launch post linking to current pricing, put “There is no free plan” in plain text on the pricing page, update the review-platform profile, and email the listicle’s editor with the pricing page link.
What I would expect, not what is guaranteed: the AI Mode errors fall once the launch post and pricing page are recrawled; the Perplexity errors persist until the listicle changes or a fresher source displaces it; the no-search ChatGPT errors persist until a later model. Three layers, three timelines, one fact.
Results that should change where you spend effort
- Errors fall only where a fixed source was cited: the lever works. Keep fixing sources, in order of how often your log shows them cited.
- Errors persist in runs where no search ran: this is memory. Stop editing pages for that platform, keep your descriptions consistent everywhere, and re-test when it ships a new model.
- Your corrected page is cited but the answer is still wrong: synthesis. Rewrite the passage so the fact stands in one explicit sentence that makes sense without the surrounding paragraphs, and check that no other page on your site contradicts it.
- The wrong claim keeps citing a third party that will not change: put effort into current, better sources that answer the same question, and expect a slow fade rather than a switch.
Where I would start
Spend the first week sampling prompts and classifying errors by layer. Fix your own stale pages next, because they are the only source you can change the same day. Send correction requests to the third-party sources that appear in citations, fix Wikidata only for company-level facts, and treat errors from model memory as a slow, aggregate problem you measure quarterly rather than chase weekly. If the error rate on retrieval-based answers does not fall after the corrected pages are recrawled, the likelier explanation is synthesis or a source you have not found yet, not proof that nothing works.
Need help with this on your own site?
Niraj Raut works with businesses in Nepal, Australia, the UK, Europe and the US on technical SEO, ecommerce SEO, local SEO and AI search.
Add a reply Cancel reply
Replies are open to approved members. Log in or apply to join.