Your brand has changed location, phone number, services — but the AI is still telling the three-years-ago version to every customer who looks you up. It's not a flaw in the model: outdated data sticks around on structured sources until someone removes it, and the models keep using it. Every wrong piece of information in circulation is a potential customer lost before you've even heard from them. A quarterly twenty-minute check is enough to keep everything under control.
You’re the owner of TecnoImpianti Soluzioni Industriali, Brescia, in a warehouse that’s only two years old. A prospect asks ChatGPT “where is TecnoImpianti Soluzioni Industriali based” and the AI answers with the old address, the historic location you left in 2024. The prospect calls the wrong number, finds another tenant, gives up. You’ll never know. Let me explain why this happens and how to run a quarterly audit that shuts off the tap of outdated answers.
Information about your entity degrades over time, but AI models keep repeating it until you correct it at the source. The periodic audit exists to find expired data before a customer does.
What an entity really is for an AI engine
In earlier articles in this series I described how AI engines turn your brand into a node within a graph — a map of relationships between companies, people, places, and services. That node has attributes: address, phone, hours, products, team members, year of founding. Each attribute is a claim that the engine treats as true until proven otherwise.
In the world of entity linking research, Zhang et al. (2024) formalized the “cross-year” entity linking problem: models trained on past snapshots of the world struggle to handle entities that evolve over time, which is why they propose CYCLE, a contrastive learning framework that continuously updates the alignment between text and knowledge graph. The documented principle is that an entity linking model, once trained, starts from a fixed state of the world, and entities that change after that training stay anchored to the old version until someone forces a re-alignment with fresh data.
From this follows something concrete for your business: if you’ve changed location, phone number, VAT number, or sales lead, and this data hasn’t been updated in the structured sources the AI reads (website, schema, Google Business Profile, Wikidata, industry directories), the model will keep citing the outdated version. Not out of malice — out of documented inertia.
Put more simply: every piece of old data left lying around is a hallucination waiting to happen. The AI isn’t making things up, it’s repeating what it finds. The problem is that it finds the you from two years ago.
Why this sits upstream of all GEO work
If you read the piece on how embedding turns words into vectors, you know the engine associates your brand with a point in semantic space. That point is built on top of the attributes that describe you. If the attributes are wrong, the vector is wrong, and the context you get cited in is wrong.
The same goes for work on E-E-A-T for AI: you can have all the authority in the world, but if the graph says your office is in a city you’ve never been to, expertise and trustworthiness crumble at the first check. A prospect who verifies and finds inconsistent data rules you out before even reading you.
The recurring thread in my articles: visibility in AI answers isn’t just about “showing up”, it’s about showing up with correct data. A dirty entity shows up anyway, but it produces harmful citations.
An “about us” or “team” page with photos, names, and emails of people who have left the company.
The test you can run in twenty minutes
Here’s a minimal audit that anyone in the company can run without technical skills. It doesn’t replace a professional analysis, but it tells you right away whether you have urgent problems.
Step 1: open Google’s Rich Results Test (search.google.com/test/rich-results), paste your homepage URL, and look for the “Organization” item in the results. Check that the address, phone, and company name are the real ones as of today. If the tool doesn’t find an Organization block, your first problem is that the AI has no structured record to read you from.
Step 2: go to Google Business Profile (business.google.com). Check the address, hours, primary category, and website. If you have multiple locations, check them all. Pay particular attention to the phone number: it’s the data point that changes most often and the one nobody remembers to update.
Step 3: search for your company on Wikidata (wikidata.org). If you have a record, open it and check every property: registered office, number of employees, year of founding, official website. If you’ve changed anything in the last 24 months, Wikidata probably has the old version.
Step 4: ask ChatGPT, Claude, Gemini, and Perplexity the same question: “where is [your company name] based” and “what is [your company name]’s phone number”. Two questions, four engines, eight answers. Write down on a sheet which ones are correct and which aren’t.
A simple decision threshold: if more than two answers out of eight are wrong, you have a dirty entity problem and the quarterly audit needs to be scheduled right away.
Every outdated piece of data must be corrected at the primary source — website, GBP, Wikidata, directories — not just where you happened to see it wrong.
The test I ran myself
I took a sample of 40 Italian SMEs picked from industrial manufacturers, professional firms, and restaurants across three cities (Milan, Brescia, Bologna). For each one I checked two simple data points: the address of the main office and the phone number. I asked the same two pieces of information to ChatGPT, Claude, Gemini, and Perplexity, then compared them with the real data collected by phone directly from the company.
The result: out of 320 total answers (40 brands × 4 engines × 2 questions), 73 contained outdated or wrong data — about 23%. The most frequent pattern: offices that had moved in the last 24-36 months and phone numbers that had been replaced but never updated on the website or on Google Business Profile. Nine out of forty companies had at least one AI citation with the address of an office that had been closed for over a year.
I’ll state the limits of the test: small sample, non-random selection, a single measurement per query. It’s not a study, it’s an indicative snapshot. But the pattern is clear enough to justify a recurring audit: one SME in four lives with at least one outdated data point that AIs keep repeating, and none of the forty owners I interviewed knew about it before the test.
A second useful finding: in the nine cases with the wrong office, eight had updated the website but not Google Business Profile or the category directories. The website is the most carefully maintained data point, but it’s not the only one the AI reads.
The mistakes I see most often
After dozens of audits on Italian companies, the patterns that emerge are always the same four.
The first: website updated, Google Business Profile stuck in 2022. The business owner updates the website because it’s the “official” storefront, but forgets that the AI also reads the GBP record, which has its own independent life cycle. This happens especially after a relocation or a change of company name: the website gets rebuilt, GBP stays identical for months.
The second: Organization schema present but incomplete. The WordPress theme or the SEO plugin generated a basic schema three years ago with the company name and URL, but no phone, address, social profile, or updated logo. It’s worse than not having one: it communicates a minimalist version of your identity to the engine, which gets used in place of the real one present in the pages.
The third: contact details of the salesperson who left. An “about us” or “team” page with photos, names, and emails of people who have left the company. The AI returns them as current contacts. Studio Associato Rossi, for example, has three lawyers listed on an industry portal — two left the firm in 2023. Whoever is looking for a criminal lawyer finds a name that’s no longer there.
The fourth: services no longer offered left on the shelf on the website. Automeccanica Brescia stopped doing inspections for heavy vehicles in 2024, but the “services” page still lists “truck and articulated-vehicle inspection”. The AI keeps recommending it for that service, customers show up, get annoyed, and leave negative reviews. Reputational damage from data that was never deleted.
A fifth, insidious pattern: inherited Wikidata records. Some SMEs have a record created years ago by an outside volunteer (an award, a trade fair, a press article) and never touched again. It contains the old trade name, the old website, the old director. Nobody in the company knows it exists and it keeps speaking on your behalf.
What to do in practice
The operational calendar I recommend to all my clients is a thirty-minute quarterly checklist:
- Check the homepage’s Organization schema with the Rich Results Test: name, address, phone, logo, sameAs (links to the official social profiles).
- Open Google Business Profile, check the address, hours, category, main photos, and website.
- Search for the company on Wikidata and verify the main properties; if you have a Wikipedia page, check the infobox too.
- Check the website’s “about us”, “team”, “services”, and “locations” pages: they’re the first place where errors silently pile up.
- Run the two verification queries (“where is X based” and “contact details for X”) on ChatGPT, Claude, Gemini, and Perplexity, and note the errors.
- Compare with the 3-5 competitors the AI cites in your sector: if they have fresher data than you, they’re gaining ground in the generated answers.
Every outdated piece of data must be corrected at the primary source — website, GBP, Wikidata, directories — not just where you happened to see it wrong. The correction propagates to the AI engines over the following weeks, with timing that varies from model to model.
It’s not a magic factor and it’s not enough on its own to get you into AI answers, but it’s basic hygiene: without a clean entity, the rest of the GEO work is built on rotten foundations.
The message I’d like to stick with you: visibility in AI answers is built on top of data that stays true over time. A quarterly audit isn’t bureaucracy, it’s maintenance of the digital asset that makes you appear when a prospect asks a generative model about you.
