Hey there 👋

Last week I said accuracy is the first thing financial brands should fix on their own pages. This week is about the other half of that problem, and that’s what AI search says about you when you're not the one writing it.

AI search gets brand facts wrong…a lot!

It quotes fees you changed last year and describes products you shut down long ago. That can be the difference between a prospect booking a demo and a prospect deciding you're too expensive.

The useful part is that most of these errors aren't random. A lot of them trace back to a real page somewhere, often one you published yourself and forgot about. That makes them fixable.

This week, I'll walk through a six-step audit: find the wrong claims, trace each one to its source, and fix the source rather than the symptom.

Let's go. 🚀

Alongside this newsletter, I help financial, fintech, and crypto brands get found on Google, ChatGPT, Perplexity, and Google AI Mode, either directly or through my agency, Rise Up Media.

If you'd like help with yours, tell me what you're working on. It takes two minutes.

Work with me →

Fix the source and the chatbots follow. Full Fact asked ChatGPT, Claude, Gemini, Grok and DeepSeek the same 22 questions every day for 70 days, with web search switched on, and checked more than 74,000 individual facts in their answers. One recurring failure was stale information, such as a UK asylum support rate that had changed in June 2025 still being quoted at the old figure. The finding that matters for brands: when the House of Commons Library published an error, it spread into chatbot answers, then largely disappeared once the Library corrected it. [Source: Full Fact]

A shared name can cost you AI recommendations. Search Engine Land looked at the UK activewear brand TALA. AI models recognized it when asked about it directly, but it didn't show up when people asked for recommendations in its category. Google's Knowledge Graph, the database Google uses to understand which entity a name refers to, listed a different company called TALA as a financial services firm. If your brand shares a name with someone else, check which one AI thinks you are. [Source: Search Engine Land]

AI-written articles are turning errors into "sources." In a Forbes Agency Council post, Xpand Digital founder Joel House describes a Google AI summary that gave creator Alberta Devor a childhood in Mexico. She's from Brooklyn. AI-written articles then repeated the claim, and those articles became the citations for the summary. His advice is to check whether the pages backing up a wrong claim are actually independent, or just copying each other. [Source: Forbes]

By the end of this, you'll have a list of every false claim AI search makes about your brand, with a likely source and a fix next to each one. There are six steps, and the order matters, because the later steps only work if the early ones are thorough.

The best evidence for how this works comes from Seer Interactive, which ran the audit on itself. Between May 1 and June 9, Seer sent 1,562 prompts about its own agency to AI Mode, AI Overviews, ChatGPT, Claude, Gemini, and Perplexity, and checked 28,123 answers. When the platforms answered, they were right 95.9% of the time. But they declined to answer 31.2% of the time, and they cited more than 9,000 different websites along the way.

Keep in mind this is one agency auditing itself, so treat the numbers as a direction rather than a benchmark. The causes it found are what's useful.

You can't audit what you don't ask. Typing "What is [Brand]?" into ChatGPT tells you almost nothing, because your buyers don't ask that. They ask "What are [Brand]'s fees?" "Is [Brand] regulated in the UK?" and "[Brand] vs [Competitor]: which is better for a small business?"

Build a bank of 30 to 50 prompts. Cover who you are, where you're based and who runs you. Cover your products, fees, limits and the markets you serve. Then add reputation questions like "Is [Brand] safe?" or "Is [Brand] legit?", one comparison prompt for each of your three closest competitors, and a few about recent changes like launches, rebrands or discontinued products.

Weight the comparisons and recent changes most heavily. Seer's comparison prompts went unanswered 80.3% of the time, which means a buyer asking "you or them?" often gets nothing useful about you. Recent changes are where stale sources do the most damage.

Run every prompt on ChatGPT, Perplexity, Gemini, Google AI Mode, AI Overviews, Claude, and Copilot. Run each one three times, in a fresh session with memory off. Answers vary between runs, and an error that shows up once in three is still an error a buyer will see.

One answer can contain five correct facts and one wrong one. If you grade the whole answer as "mostly fine," the wrong fact survives.

Break every answer into individual factual claims and log each one on its own row:

Quote the wrong claim exactly. The precise wording is your best clue to where it came from, because answers often borrow the phrasing of the page they pulled from.

The type of error tells you where to look, and where to look tells you how to fix it. Here's a classification that makes the next step easier:

Only the last row is a true hallucination. Seer found the same split: some errors came from training data, but others came from the model pulling a stale page from the web, or misreading the brand's own site. If you assume every error is a hallucination, you'll do nothing. If you assume each one has a source until you've ruled that out, you'll usually find it.

Treat each wrong claim like a rumor and follow it home. Work from the fastest checks to the slowest.

Start with the citations. Perplexity, AI Mode, AI Overviews, and Copilot show their sources, and so does ChatGPT when it searches the web. That's less often than people think: Ahrefs found only 14.4% of free, logged-out ChatGPT answers triggered a search in August and September. Open every cited link and search the page for the wrong claim.

If there are no citations, search Google for the most distinctive part of the claim in quotation marks, plus your brand name, for example, "0.5% fee on card top-ups." Then search the wrong number on its own.

Next, check your own site. Seer's own website made up more than 35% of the citations in its study, so this is where most stale facts live. Run a site: search on Google for your domain plus the outdated fact, and go through old blog posts, archived fee pages, PDF factsheets, help center articles and forgotten subdomains. The Wayback Machine shows what your pages used to say.

Then sweep your third-party profiles: Crunchbase, LinkedIn, Trustpilot, app store listings, Wikipedia and Wikidata, Google Business Profile, comparison sites and industry directories. Seer found about a dozen sources it could actually influence, including LinkedIn, Clutch and Reddit. These rarely get updated after they're first set up.

Finally, ask the same question with web search on and off, on platforms that let you switch. If the error only appears with search on, a live page is feeding it. If it appears with search off, it's coming from the model's training data.

Don't ask the model where it got a claim. If it didn't search, it has no way of knowing what it was trained on, and you'll get a confident, invented answer. The only sources worth trusting are the links it actually showed you.

Fixing the page an error came from stops it from spreading. Publishing a clear answer stops a new one from taking its place.

Match the fix to where the error came from. On your own site, update or remove the outdated page, redirect retired product and fee pages to the current ones, and replace old PDFs. Add organization and product schema that matches what's on the page. It helps engines understand which entity you are, even though it isn't a citation booster.

For third-party profiles, claim every listing you don't control yet and update it, then set a quarterly reminder to check them again. For news coverage that's factually wrong, ask for a correction. If it was accurate at the time and is now just old, publish a fresh announcement so there's a newer source to prefer.

Competitor and affiliate comparisons need a correction request with evidence, followed by your own fair, specific comparison page. If you don't explain how you compare, someone else will, and Seer's 80.3% non-answer rate on comparisons shows how often nobody has.

For forums and review sites, reply publicly from an official account with the current facts, so the correction sits on the same page as the error. Gaps and hallucinations need the missing fact published plainly. A facts page with your founding year, headquarters, regulators, fee structure, and markets served gives models a clean answer instead of a hole to fill. Use each platform's feedback button on wrong answers too.

Write every correction as a plain, quotable sentence. "[Brand] charges a 0.5% fee on card top-ups over €1,000" is far more useful to an AI engine than "transparent pricing for everyone." Seer found misreadings of its own site were the fastest errors to fix, usually by making the wording more explicit.

Re-run your prompt bank every month and track three numbers: the share of claims that are wrong, the number of open errors by type, and how long each error takes to disappear after you fix it.

Expect two speeds. Seer found errors from live web pages can clear within days or weeks once the source is updated, while errors baked into training data take months and depend on the model being retrained.

Re-test outside the monthly cycle after any big change on your side, such as a fee change, rebrand, acquisition, new license or product shutdown. That's when AI search is most likely to fall behind.

After your first full audit, you should have:

This turns a pile of copied AI answers into a clean error log, so step 2 takes an hour instead of a day.

The Scenario: You've run your prompt bank and pasted the answers into a document. Now you need a claim-by-claim error log with suspected sources, without doing it all by hand.

The Prompt:

Run it in a fresh chat for each batch of answers, and keep batches small enough that you can check the table yourself.

These are the most popular AI tools we use at Rise Up Media. If you're not using them already, they're worth a look.

Full disclosure: some links above are affiliate links. If you sign up, we'll earn a small commission at no extra cost to you.

Ask the questions your buyers ask, log every claim, trace the wrong ones to a page, and fix that page before you worry about the model.

Most of what AI gets wrong about a brand was written somewhere first, often by the brand itself, years ago.

Full Fact watched a single correction from the House of Commons Library pull an error back out of chatbot answers. Your old fee page works the same way.

Until next time, keep exploring the horizon. 🌅

Alex Lielacher

P.S. If you want your brand to show up in Google AI Mode, ChatGPT, and Perplexity, reach out to my agency, Rise Up Media. That's what we do!