To audit what language models say about your brand, ask several assistants a fixed set of questions about your business, record the answers, mark every error and gap, trace each error to the source it probably came from, and fix it there. You can't edit a model directly. You can change the sources it learns from and retrieves, and with web-connected assistants the corrections often show up within weeks.
Why errors happen
Models get businesses wrong for a few predictable reasons:
- Name collisions. Another company, product or person shares your name.
- Stale sources. An old directory listing, a former address, a service you dropped years ago.
- Thin evidence. Too few independent sources, so the model fills gaps with plausible guesses.
- Inconsistent self-description. Your own website, LinkedIn and Google profile describe you differently.
- Training lag. The model's training data predates a rebrand or a new service.
Step 1: build the question set
Write 15 to 25 prompts in four groups:
- Identity: "What is {brand}?", "Where is {brand} based?", "Who founded {brand}?"
- Offer: "What services does {brand} offer?", "How much does {brand} charge?"
- Category: "Best {service} agencies in {city}", "Who does {service} in South Africa?"
- Comparison: "{brand} vs {competitor}", "Is {brand} good for {use case}?"
Keep the wording fixed so you can compare month to month.
Step 2: run it across assistants
Run every prompt in ChatGPT, Claude, Gemini, Perplexity, Copilot and Google AI Mode. Where you can, note whether the assistant searched the web (most show sources when they do) or answered from its own knowledge. That distinction tells you where to fix things:
- Answers with sources: the error is in a page the assistant retrieved. Fix or outrank that page.
- Answers without sources: the error is in what the model learned. Fixing it takes longer and relies on consistent sources over time.
Do each run in a fresh session, because chat history and personalisation change answers.
Step 3: score the answers
For each answer, record:
- Accuracy: correct, partly wrong, wrong.
- Completeness: are your main services mentioned?
- Sentiment and positioning: how are you described relative to competitors?
- Sources cited: which URLs, if any.
A simple spreadsheet works. After one run you'll usually see a pattern, such as one wrong fact repeated everywhere.
Step 4: trace each error
For each wrong fact, search for the exact wording. Common culprits:
- An old Google Business Profile or Bing Places listing.
- A directory that scraped your details years ago.
- A press article that described you loosely.
- A similarly named company abroad.
- Your own outdated about page or a forgotten landing page.
Step 5: fix at the source
- Your own site: update the about page, contact page and homepage. Make the correct facts explicit ("Arkces is based in South Africa and was founded by Ira Emery") and back them with Organization structured data. Our post on entity SEO has the details.
- Profiles you control: correct them all to match one canonical description.
- Third-party sites: ask for corrections. Most directories and publications fix factual errors when asked politely with evidence.
- Name collisions: add disambiguating detail everywhere ("Arkces, the South African digital agency"), and use the same descriptor consistently.
- Thin evidence: earn more independent mentions with the right facts. See brand mentions vs backlinks.
Some assistants also have feedback buttons on answers. Use them for clear factual errors, but treat them as a supplement, not a fix.
Step 6: re-test and keep a log
Repeat the full set monthly. Web-connected answers often improve within a few weeks once the source is fixed and recrawled. Answers from model knowledge may take until the next model update. Keep a log of each error, its source, the fix and the date, so you can tell what worked.
Make sure crawlers can see the fix
Corrections only help retrieval if AI search crawlers can read your pages. Check robots.txt and your CDN for OAI-SearchBot, Claude-SearchBot and PerplexityBot. OpenAI and Perplexity both note that changes to robots.txt take around a day to register.
When to get help
If the errors are damaging, such as a wrong location that loses you local leads or a competitor's reviews attached to your name, it's worth a structured programme. This audit-and-correct loop is the core of our LLM SEO service.
