Why AI assistants describe your brand wrong, and how to fix it

Updated

Last checked: Nov 3, 2026. The facts about how ChatGPT and Perplexity work come from the vendors' own help and documentation pages, checked on Nov 3, 2026 (sources at the end). The numbers come from our own daily tests of what AI assistants say about MentionSweep, and we say exactly how we ran them.

Availability (Nov 3, 2026): MentionSweep's free score, $29 Full Report, $39/month Monitor and Agency plan are all live, with checkout through Stripe.

The short answer

AI assistants get small brands wrong for two different reasons, and each needs a different fix:

  1. Answering from memory. Without web search, a model only knows what was in its training data. A new or small brand often isn't there, and some models fill the gap with a confident guess based on the name. You can't edit that memory. In our tests, these answers didn't change in 15 days.
  2. Answering from the web, from the wrong page. With web search, an assistant repeats whatever page it finds about you. If that page is a third-party listing with old details, you get old details back. This one you can fix, by making sure the pages it finds are clear and correct.

So the work is: give search-backed assistants one clear, correct page to find, fix the listings they cite, and check again on several days.

Why it happens

Memory answers: a guess from the name

OpenAI's help center is direct about this. Without search, ChatGPT's "responses are based on what the model has learned during training", the models are trained on data "up to a certain point", and a model "may express high confidence even in incorrect answers". The same page lists "incorrect definitions, dates, or facts" as one kind of hallucination (Does ChatGPT tell the truth?).

For a brand the model has never seen, that's a recipe for invention. Our name is a good example: "MentionSweep" reads like "mention" + "sweep", and the guesses follow the words, not the product.

Search answers: only as good as the page found

With search, the assistant looks you up first. OpenAI's help center says "search results and citations can be incomplete, outdated, or incorrect" and tells users to open the cited source and check it (Searching the web with ChatGPT). From the brand's side, that means the cited page is your description. If it's out of date, so is the answer.

Whether search is used depends on the product and the settings. ChatGPT "may search the web automatically" when a question would benefit from current information (same page). In Perplexity's Agent API, "web search is opt-in": a bare request "answers from the model's own knowledge and returns no citations" (Perplexity Agent API docs). So the same assistant can give a memory answer one time and a search answer the next.

What we saw when we asked about ourselves

Since Oct 19, 2026 we've asked "What is MentionSweep?" every day through the AI providers' official APIs, along with 7 buyer questions about AI visibility tools. This page uses 15 run days: Oct 19 to 24 and Oct 26 to Nov 3 (our Oct 25 run didn't happen because of an outage on our side). Engines were added as we got access: ChatGPT and Gemini from Oct 19, Perplexity from Oct 20, Claude from Oct 21, Grok from Oct 22.

How we asked: ChatGPT through the OpenAI API (gpt-5-mini), once without web search and once with OpenAI's web search tool; Gemini without Google Search grounding; Claude (claude-haiku-5-5) and Grok (grok-4.20-0309-non-reasoning) without web search; Perplexity through its Agent API, which returned web citations in our runs. These are API answers, not the ChatGPT app or any signed-in app, so what you see in an app can differ.

Gaps, said plainly: Perplexity couldn't be checked from Oct 30 to Nov 3, because our Perplexity API requests were refused for billing reasons, and ChatGPT without web search couldn't be checked on Oct 31 because of network errors on our side. Those 6 are left out, not counted as "no".

"What is MentionSweep?": 78 answers over 15 days

Assistant and mode Answers Correct Made up a different product Said it didn't know, or couldn't find us
ChatGPT, no web search 14 0 3 11
Gemini, no web search 15 0 15 0
Claude, no web search 13 0 0 13
Grok, no web search 12 0 12 0
ChatGPT with web search 15 3 0 11 (+1 mixed)
Perplexity (web search) 9 5 0 4
All 78 8 30 39 (+1 mixed)

Five things stand out.

1. From memory: 0 correct in 54 answers, 30 made up. Gemini described MentionSweep as something else on all 15 days, usually a social-listening tool that watches Reddit and X for keywords, and on Nov 3 a tool for mass-deleting old mentions on X. Grok made something up on all 12 of its days: most often an OSINT tool that finds accounts mentioning a username, other days an open-source script that deletes old X mentions. ChatGPT without search did it on 3 days and said it didn't know on the other 11. MentionSweep is none of these: it checks whether AI assistants name your brand.

2. "I don't know" is possible. Claude said it didn't know in all 13 of its answers. On Nov 3 it offered a guess from the name and then added "I'd only be guessing". That's the honest answer for a brand a model hasn't seen, and it shows the made-up answers aren't inevitable.

3. Fixing our own site didn't change the memory answers. Since Oct 24, our About page has said in plain words that MentionSweep is "not a social-listening, media-monitoring, OSINT or tweet-deletion tool". In the 9 run days after that (Oct 26 to Nov 3), Gemini still made up a product 9 times out of 9, and Grok 9 out of 9. A model answering from memory doesn't read your site.

4. Every correct answer came from a web search, and most came from one listing. All 8 correct answers were from assistants that searched the web. 7 of the 8 led with the same directory listing about us (on geodeck.io) as their first or only source, not our own site. Only one, Perplexity on Oct 29, led with our own About and home pages (it listed the directory last). Before that listing went live on Oct 22, the search-backed assistants answered "not found". On Nov 3, ChatGPT with web search put it plainly: it "couldn't find an official product page, but directory listings describe MentionSweep as an AI-visibility / answer-engine monitoring tool".

5. The listing's details became the answer's details. The listing describes our free score as running "across ChatGPT, Gemini, Claude, Grok and Perplexity". Our free score asks four assistants; Perplexity is only in the paid report. All 7 answers that led with the listing repeated the five-name list without that split, including ChatGPT with web search on Nov 2 and Nov 3. The one answer that led with our own pages (Perplexity, Oct 29) got it right: "free check covers ChatGPT, Gemini, Claude, and Grok; paid reports and monitoring add Perplexity".

It also doesn't stick. ChatGPT with web search was right on Oct 29, couldn't find us on Oct 30, mixed us up with a differently named "brand mention sweep" tool on Oct 31, and was right again on Nov 2 and Nov 3. Perplexity was right on Oct 23 and 24, not on Oct 26, then right again Oct 27 to 29.

For buyer questions without our name in them ("what's the best AI visibility tracker?" and 6 others), none of the 545 answers over the same 15 days named MentionSweep. Being described correctly when someone asks about you comes first; being recommended when they don't is a separate, slower problem. See which AI visibility tools AI assistants recommend for who does get named.

How to fix what AI says about your brand

1. Find out which kind of wrong you have

Ask each assistant "What is [your brand]?" and note whether the answer shows sources. No sources and a confident wrong description: a memory answer. Sources and a wrong detail: open the cited page, because that's where the detail came from. Our guide on how to check what AI says about your brand walks through it step by step.

2. Publish one page that says plainly what you are, and what you're not

Search-backed assistants need something to find. Write a short "About [brand]" page in plain text: one sentence on what you are and who it's for, what it costs, when it launched, your official website, and, if your name invites a wrong guess, a line on what you're not. Put real numbers in the text, not only in images or scripts. Our own About page is an example.

3. Make that page findable

A clear page only helps if the assistant's search can reach it. OpenAI says that to be eligible for ChatGPT search results, a site should allow OAI-SearchBot and make sure its host or CDN allows traffic from OpenAI's published search-bot IP addresses (Searching the web with ChatGPT). Get indexed in regular search engines too. Our guide to getting recommended by ChatGPT covers crawlers, indexing and IndexNow.

Even then, expect a lag. Our About page has been live since Oct 20, and on Nov 3 ChatGPT with web search still said it couldn't find an official product page.

4. Fix the third-party listings first

This was the biggest lesson in our data. A directory listing can reach search-backed assistants faster than your own site, and when it's out of date, they repeat it. List every directory, review site and roundup that describes you, check each one against your current product, and send short, factual corrections when something is out of date. Check again after the listing updates, because a correction you sent isn't a correction that's live.

5. Keep every fact the same everywhere

If your homepage, pricing page, About page and listings each describe your plans slightly differently, an assistant may quote any of them. Use the same plan names, prices and feature lists everywhere, and update them all on the same day when something changes.

6. Don't expect memory answers to change soon

You can't send a correction to a model's training data. What a model knows from memory changes when the provider trains a new model, which you don't control. Until then, the realistic goals are: assistants that search find a correct page about you, and you know which memory answers are wrong so you can address them on your own site, in sales and in support.

7. Re-check on several days, not once

AI answers about a small brand change from day to day, as ours did. Keep a fixed question list, record the answer, whether it had sources and which pages were cited, and look at the trend over a week or more.

Checklist

  • Asked "What is [brand]?" in each assistant and noted which answers had sources
  • Opened every cited page and checked it against your current product
  • Published a plain "About" page: what you are, who it's for, price, website, and what you're not
  • Search crawlers allowed in robots.txt and at the CDN; site submitted to search engines
  • Listed every directory and review page that describes you, and sent corrections where needed
  • Same plan names, prices and features on every page and listing
  • Re-check scheduled for next week, with the same questions

Ways to check

You can do all of this by hand in a spreadsheet. MentionSweep's free score asks ChatGPT, Gemini, Claude and Grok 10 buyer-style questions written from what you sell, through their APIs without web search, and gives you a 0 to 100 score in about 30 seconds. The $29 Full Report is a one-time check that adds Perplexity, with share of voice against up to 3 competitors and the sites AI cites. Monitor ($39/month) re-runs the questions every week. For other tools, see our comparison of free AI visibility checkers.

MentionSweep isn't affiliated with OpenAI, Google, Anthropic, xAI, Perplexity or any vendor named here, and no tool can make an AI assistant describe or recommend you.

Sources

  • OpenAI Help Center: Does ChatGPT tell the truth? (checked Nov 3, 2026). Hallucinations include "incorrect definitions, dates, or facts"; knowledge cutoff; "confidence isn't reliability"; without search, responses are based on what the model learned in training.
  • OpenAI Help Center: Searching the web with ChatGPT (checked Nov 3, 2026). ChatGPT may search automatically; "search results and citations can be incomplete, outdated, or incorrect"; to be eligible for ChatGPT search, allow OAI-SearchBot and OpenAI's search-bot IP addresses.
  • Perplexity docs: Agent API (checked Nov 3, 2026). "Web search is opt-in"; without it, answers come from the model's own knowledge with no citations.
  • Geodeck: AI visibility monitoring directory and its MentionSweep listing (checked Nov 3, 2026), the main source in 7 of our 8 correct answers.
  • MentionSweep: our daily API checks, Oct 19 to Nov 3, 2026 (no Oct 25 run); our About page; free score and plan details from our own product.

Want to see what ChatGPT, Gemini, Claude and Grok say when your buyers ask? Get your free AI visibility score.

FAQ

Questions, answered

Why does ChatGPT or Gemini describe my brand as something it isn't?

Without web search, an AI model answers from what it learned in training. OpenAI's help center says that without search, ChatGPT's responses are based on what the model learned during training, up to a knowledge cutoff, and that it can sound confident even when it's wrong. If your brand is new or small, the model may never have seen it, and some models then guess from the name. In our own tests over 15 days, AI models answering from memory made up a different product in 30 of 54 answers about MentionSweep.

Can I make an AI assistant correct what it says about my brand?

Not directly, and not from memory. What you can change is what an assistant finds when it searches the web: your own clear 'about' page, consistent facts across your site, and accurate third-party listings. In our tests, every correct answer about MentionSweep came from an assistant that had searched the web and cited a page about us.

How long does it take for AI answers about my brand to change?

For answers that use web search, it can be days once a clear page about you is findable: our first correct answer came one day after a directory listing about us went live. For answers from memory, we saw no change at all in 15 days, including in the 9 run days after we added a plain 'what MentionSweep is not' section to our About page.

Why does a third-party listing matter so much?

Because search-backed assistants cite it. In our data, 7 of the 8 correct answers about MentionSweep led with the same directory listing as their main source, not our own site. When that listing's details were out of date, the assistants repeated them.

Does MentionSweep check what AI says about my brand?

Yes. The free score asks ChatGPT, Gemini, Claude and Grok 10 buyer-style questions through their APIs and gives a 0 to 100 score. The $29 Full Report is a one-time check that adds Perplexity. Monitor ($39/month) re-checks every week.

Is ChatGPT recommending your competitor?

Check in about 30 seconds. Free, no card.

Get my free AI score