How do you fix what AI assistants say about your company? There is no correction form for a language model. You change what it reads. Measure the current answers, fix the surfaces assistants retrieve live, correct the third-party sources they cite, publish content that answers the buyer’s problem question, and re-measure at every model release. Source fixes show up in weeks. Trained knowledge catches up on the next model.
Your company changed its story. Ask ChatGPT what your company does and there is a meaningful chance it answers with the story you retired. I audited 47 B2B SaaS companies that renamed, rebranded, or changed category between January 2024 and October 2025. A median of 21 months after their announcements, 87% still had at least one major public surface telling the old story, and when I asked a frontier model what each company does, one in seven was described purely as the thing it had spent millions to stop being.
This piece is the remediation sequence. It draws on 9,774 model runs across three published waves of SandsDX research into how AI assistants describe B2B software vendors, and on running this fix for clients. It assumes you have already repositioned and the assistants have not caught up.
Why the model says the wrong thing
An AI assistant builds its answer from two places. Trained knowledge: what the model absorbed about your company during training, frozen at a cutoff that likely predates your change. Live retrieval: the pages the assistant reads at answer time, review sites, directories, comparison articles, Wikipedia, your own site.
Both layers hold the old story, and they fail differently. Trained knowledge fails by omission. The model that shipped before your announcement cannot know about it. Retrieval fails by consensus. The assistant reads five sources, four still describe the old category, and the majority wins. My audit found the second failure is the durable one: the median company had multiple stale surfaces feeding every retrieval pass, 21 months in.
There is a third, stranger failure. Wave 1 of the AI Shortlist Report measured a 14-point gap between models agreeing on which companies belong in a category (79%) and agreeing on what those companies are called (65%). The models frequently know a rebrand happened and default to the old name anyway, because the old name dominates the text they learned from. Recognition arrives before adoption. Your job is to shorten the distance between the two.
What does not work
Waiting. The 21-month median is what waiting looks like, and most of the 47 companies were still waiting when the audit closed.
Petitioning the lab. None of the major AI providers accepts corrections to model knowledge about a company. There is no form, no representative, no expedited review.
A press release. Announcement coverage decays in days. The surfaces that feed retrieval, review platforms, directories, comparison posts, persist for years and outvote your newsroom.
The remediation sequence
1. Measure the baseline before touching anything
Run the questions your buying committee actually asks, across the assistants your buyers use, and record where you appear, under which name, in which category, and who holds the places you do not.
One generic prompt reads almost nothing. Wave 3 ran five kinds of buyer questions across 26 categories and 8 frontier models: stating the buyer’s company size changed the rankings four times more than rewording the question, and describing the problem instead of naming the category produced a different leaderboard entirely. Ask by size band, by problem statement, by head-to-head comparison, and by category term. The AI Visibility Baseline runs the persona-level version of this and is free.
While measuring, ask each assistant to cite its sources. That list is your work queue, ordered by evidence.
2. Fix the surfaces you control
Your own site is the one retrieval source you can correct today. The checklist for a repositioned company:
- New name and new category in title tags, H1s, and meta descriptions, sitewide, with no legacy stragglers on old blog posts and case studies.
- Organization schema with the current name, the current category description, and
sameAslinks to every profile that confirms the new story. Run the Structured Data Detector to see what your pages currently declare. - An
llms.txtfile that states the current positioning in plain language. Assistants that read it get your story in your words. - A short bridge statement where it helps: “formerly known as X” or “previously positioned as Y,” stated once, prominently. The naming-gap data shows models need the old-to-new association spelled out. Pretending the old name never existed slows adoption of the new one.
- Extractable definitions near the top of key pages. Assistants quote what they can lift cleanly.
3. Correct the sources assistants cite
Work the queue from step 1 in order of citation frequency. In practice the queue concentrates in five places:
Wikipedia and Wikidata. Heavily weighted, publicly editable, governed by sourcing rules. Update requires secondary coverage of the change, which your announcement coverage provides. This is the single highest-yield edit for companies notable enough to have a page.
Review platforms. G2, Capterra, TrustRadius. Category placement, product descriptions, and comparison grids are vendor-editable through their programs. In the audit, review-site category pages were among the most common stale surfaces.
Directories and analyst listings. Crunchbase, vendor directories, market landscape pages. Individually small, collectively they form the consensus that outvotes your website.
Comparison content. The “X vs Y” and “best tools for Z” posts ranking in your category feed retrieval directly. The authors update for accuracy more often than vendors assume, and an accurate-correction request with a source succeeds more often than a pitch.
Press boilerplate. The description that wire services and journalists copy forward. Fix the master version; it propagates.
4. Publish the problem-query answer
Wave 3’s sharpest finding: ask assistants to name vendors by category and one set of companies wins. Describe the buyer’s problem in plain language and a different set wins. Buyers ask the problem question first, before they know what the category is called, and most complete their evaluation before ever contacting sales.
A repositioned company has usually written extensively about its new category and almost nothing that answers the problem in the buyer’s words. Write the problem pages: a direct definition the assistant can lift, the specific pains, named alternatives, and numbers with sources. Statistics and citations measurably increase the odds of being quoted in AI answers. This is the same discipline covered in the GEO/SEO executive guide, pointed at the question your buyer asks on day one.
5. Re-measure at every model release
A model release is a market event for your visibility. Wave 1 measured a single release: 13 of 21 category shortlists changed, 13% of all top-five places changed hands, with no market event in between. Your remediation can be undone, or accelerated, by a training run you had no part in.
Re-run the step 1 baseline within two weeks of every major release from the labs your buyers use, and quarterly regardless. Log rank, naming, and cited sources each time. The trend across runs, not any single reading, tells you whether the new story is taking.
What to expect
Retrieval moves first. Fix your site, Wikipedia, and the top cited sources, and answers that browse can change inside a month. Trained knowledge moves on release cycles: the model that learned your new story arrives with the next training run, which is why step 5 exists. The honest framing for your board: weeks for the answers built on live sources, one to two model generations for the rest, against a do-nothing median of 21 months and counting.
Where this sits in the larger rebuild
AI answers are one surface of one system. The same repositioning that left ChatGPT behind usually left your targeting filters, your sales motion, and your attribution telling the old story too. That read across all four systems is the Gap Map, a two-week fixed-scope diagnostic with every finding cited to source.
Start with the measurement either way. The AI Visibility Baseline maps your buying committee, runs the buyer-shaped questions live, and shows you exactly which story the assistants are telling. It takes a few minutes, and it is the same baseline this sequence starts from.