Most AI-visibility reporting measures presence: how often you’re mentioned, how often you’re cited, how your share of voice compares with three competitors. Useful numbers, and they all skip the failure that actually costs money. That failure is misrepresentation: the assistant names you, cites you, and then tells a buyer you’re in the wrong city, offers a service you retired two years ago, or have a phone number that rings to a competitor’s location.
Presence without accuracy is a liability. A competitor who isn’t mentioned loses nothing. A business that is mentioned wrongly loses the buyer who believed the answer.
This guide gives you a tight, repeatable method for auditing attribute accuracy, tracing each error back to its likely source, and fixing the errors in the order that matters. It takes one focused sitting to set up and about thirty minutes a month to maintain.
Three problems that look like one
AI answer quality about a business breaks down into three separate problems, and each one needs a different fix:
- Mention: does the assistant surface you at all for relevant prompts? This is a findability and corroboration problem.
- Citation: does it link to your pages as the source? This is a content-extractability and authority problem.
- Attribute accuracy: when it describes you, are the facts right? This is an entity-evidence problem: your own pages and third-party pages disagree with each other, or say nothing where they should say something.
Teams that merge these three into one “visibility score” end up chasing mentions while the assistant keeps inventing their pricing. Separate them, and attribute accuracy becomes the leading indicator. It is also the problem you control most directly, because the evidence it depends on is largely evidence you publish.
There’s a structural reason this works. Google states plainly that there are no additional requirements or special optimizations to appear in AI Overviews or AI Mode: the same indexed pages, the same fundamentals. Answer engines summarize the web’s evidence about you. When their summary is wrong, the evidence is usually contradictory, stale, or missing, and that diagnosis is actionable.
Step 1: Build the canonical attribute record
Before you open a single chat window, write down what’s true. Pick eight attributes and record three things for each: the canonical value, the source-of-truth URL on your site where it’s stated in plain text, and the corroboration count, meaning how many independent third-party pages state it the same way.
| Attribute | Canonical value | Source-of-truth URL | Corroboration |
|---|---|---|---|
| Trade name (and legal name, if different) | /about/ | ||
| Primary location | /contact/ | ||
| Core offer (one line) | / | ||
| Who it’s for | /about/ | ||
| Differentiator | /about/ | ||
| Canonical URL | — | ||
| Preferred contact path | /contact/ | ||
| One verifiable proof point | /about/ |
The corroboration column is where the method pays off. An attribute stated on your site and nowhere else is fragile. An attribute stated identically on your site, your business profiles, and two industry directories is sturdy. You’ll use this column in Step 4 to predict where errors come from before you’ve even run a prompt.
Two rules keep the record honest. First, it reflects what’s published, not what you wish were published. If your differentiator lives only in a pitch deck, the record shows a blank, and that blank is your first finding. Second, every value must appear in text, not just inside an image or a PDF. Google’s own list of fundamentals for AI features includes “making sure that important content is available in textual form” (source).
Step 2: Design prompts in three tiers
Random prompting produces random findings. Use three fixed tiers and hold them constant from month to month:
- Identity prompts: “What does [Name] do?” / “Where is [Name] based?” These test the record directly.
- Placement prompts: “Who are good [category] providers in [city] for [audience]?” These test whether you’re retrieved, and how you’re described when you’re one option among several. Errors in placement answers are often worse than identity errors, because they sit right next to your competitors.
- Comparison prompts: “[Name] vs [Competitor] for [use case].” These test whether your differentiator survives contact with a rival’s positioning.
Run each tier in two or three assistants. Record the session type (logged out, fresh account, or your usual account) and the date. Keep the prompts boring on purpose: clever prompts measure your prompting, not the assistant’s model of your business.
Step 3: Score with six verdicts, not five
Most grids score Correct / Partial / Wrong / Missing. Add two more:
- Correct: matches the canonical value.
- Partial: right but incomplete (the city without the service area, say).
- Wrong: a real-world fact stated incorrectly, often borrowed from a similarly named entity.
- Stale: was true once. An old address, a retired service line, a former founder title.
- Invented: no basis anywhere. A phone number that never existed, a review claim nobody made, an award you didn’t win.
- Missing: the attribute never comes up.
Splitting Stale from Wrong is the most useful refinement in the grid, because the two have completely different fixes. Stale errors point to old evidence that’s still live. Wrong errors point to entity confusion. Lumping them together sends you hunting in the wrong place.
Paste the raw answer text beside each score, with the date. Never rescore an old row after you’ve fixed something. Add a new dated row instead. The history is the audit.
Step 4: Run error forensics
Every non-Correct verdict has a likely origin. This table turns a pile of wrong answers into a work queue:
| Verdict | Likely origin | First fix |
|---|---|---|
| Stale | Old directory listings, archived press, an outdated page still indexed | Update or redirect the old page; correct the listings that rank for your name |
| Wrong (conflation) | Name collision with another entity | Disambiguate: distinctive descriptors in your About copy; legalName, alternateName, sameAs in Organization markup |
| Invented | Silence: you never published the fact, so the model filled the gap | Publish the fact plainly, in text, on the source-of-truth URL |
| Partial | Facts buried in images, PDFs, or JavaScript-only widgets | Move them into HTML text |
| Missing | Weak retrieval: thin corroboration or blocked crawlers | Build corroboration; verify crawler access |
The insight hiding in that table: invented attributes cluster where you published nothing. Hallucination isn’t evenly spread. It concentrates in your evidence gaps. If an assistant invents your pricing model, the fastest fix usually isn’t a correction request. It’s a clear pricing-model sentence on a page you control.
For disambiguation, Organization structured data does real work. Google describes it as helping to “disambiguate your organization in search results” and recommends placing it on the home page or the page that describes the organization, with name and alternateName matching your site name and sameAs pointing to your profiles (Google: Organization structured data). Pair the markup with matching visible text. Google’s AI-features guidance specifically calls for structured data that “matches the visible text on the page.”
Step 5: Fix in this order
When the queue is long, sequence it by harm:
- Invented or Wrong values on money and safety attributes: phone, address, pricing, licensing. A buyer acts on these directly.
- Conflation errors: every answer about the other entity contaminates yours until you fix it.
- Stale values: usually cheap to fix and quick to propagate once the old page is gone.
- Partial values: a text-extraction pass across key pages.
- Missing: last, because it’s the slowest to move and depends on everything above.
That order is deliberate. Teams often start with Missing because it feels like growth. But extending reach while your description is still wrong just spreads the misdescription to more buyers.
Worked example: a pottery studio
Illustrative example. Cedar & Kiln Studio is a hand-built stoneware studio in Madison, Wisconsin, with weekend workshops and a small retail shop. Its first grid run, across three assistants and nine prompts, looks like this:
| Attribute | Assistant A | Assistant B | Assistant C |
|---|---|---|---|
| Location | Wrong (Milwaukee) | Correct | Stale (former studio address) |
| Core offer | Partial (no workshops) | Correct | Correct |
| Contact path | Invented (phone) | Missing | Correct |
| Differentiator | Missing | Partial | Missing |
The forensics go quickly. The Milwaukee error traces to a similarly named gallery, so it’s a conflation fix: add the neighborhood and “hand-built stoneware” to the About intro, add alternateName and sameAs. The stale address traces to a craft-fair directory page from a previous season, so the fix is to request the correction. The invented phone number appears because the site lists only a contact form, a silence the model filled. The studio adds “Contact us via the form below; we don’t take phone bookings” in plain text. The missing differentiator appears because the About page buries it in a photo caption.
Four fixes, each mapped to a specific failure, and nothing in the list is guesswork.
Step 6: Make it a monthly instrument
Re-run the same prompts on the same day each month. Track three trend lines: the Invented count, the Stale count, and the corroboration count for your weakest attribute. When Invented falls and corroboration rises, your entity evidence is converging. When a new Wrong appears, check for a new name collision before you assume a model regression.
Answer engines will keep changing models. Your dated grid is the constant, and it turns “the AI says weird things about us” into a defensible list of evidence fixes.
Move faster with the right instruments
If you want the first pass done in minutes, Organa Obscura’s free AI-Visibility Quick Check is a one-page, printable checklist of ten checks: consistent naming, a clear About page, Organization schema, crawler access for Googlebot, OAI-SearchBot, Claude-SearchBot, and PerplexityBot, answer-ready pages, freshness, third-party mentions, an AI-assistant spot check, and a dated baseline. It’s the fastest way to see which parts of this method need attention first. (More on it at organa.agency/free-ai-visibility-quick-check/.)
When you’re ready to run this as a full, repeatable program, the AI-Visibility (AEO) Audit maps directly onto the steps above. Its 27-page guide includes an AI answer spot-check method with 10 starter prompts and a worked example on a fictional business. The Sheets/Excel workbook gives you an AI Prompt Log tab for your dated raw answers, a Competitor Comparison tab for placement and comparison prompts, 60 checks scored 0/1/2 across eight audit areas, a Fix Plan that ranks work automatically by Impact × (4 − Effort), and a Summary that shows your score by area. The forensics in Step 4 slot straight into the Fix Plan, and the Quick Check is included.
Leave a Reply