AI Visibility Analyst

Why it matters

The panel is the foundation, and the questions change by business. A distributor's are part-number and cross-reference shaped: "Gates equivalent of Parker 387 hose." A roofer's are local and job shaped: "best roofer in Tampa," "metal roof cost in Phoenix." A dental group's are treatment and location shaped: "Invisalign near me," "emergency dentist downtown Austin." You collect what people type. You don't invent it.

In the field

A 60-location roofing company wanted to know how it showed up in ChatGPT. We wrote 20 prompts: a city-and-service question for each big metro, plus a few "is X roofing legit" and "metal vs shingle in Florida" ones. Run on three engines, the company surfaced in two of twenty. That low number is the panel doing its job.

Edge cases

• Vanity prompts you'd love to win but no buyer types. Drop them.
• "Near me" questions that change answer by where the asker is standing, so one run isn't the whole picture.
• A panel so big you can't run it on a schedule. Twenty sharp prompts beat 200 you never re-run.

Why it matters

The tool category is funded and crowded, and no two count the same thing. Profound and Peec lean enterprise tracking. Otterly is the cheap entry point. Scrunch and Ahrefs Brand Radar add their own angles. A distributor watching part-number queries, a dental group watching "dentist near me," and a roofer watching city prompts can all run the same tool and get numbers that don't compare. Read the definitions, not just the dashboard.

In the field

A dental group's marketer pulled an "AI visibility score" from one tool and panicked at the low number. It was counting brand mentions across the whole web, not citations (the answer engine actually linking you as a source) for treatment queries. Two tools, two definitions of the same word. The panic was about the wrong metric.

Edge cases

• A tool reporting "mentions" when the buyer thinks it means "citations." Different things.
• Free or cheap tools that sample a handful of prompts and call it coverage.
• Trusting one tool's number as truth instead of running the panel yourself to confirm it.

Why it matters

A wrong answer costs more in some businesses than others. For a distributor it's a wrong pressure rating or an interchange (the cross-reference that says part A replaces part B) stated backwards, and a customer orders the wrong part. For a dental practice it's a closed location still listed as open, or a service the practice doesn't offer. For a roofer it's the wrong service area or a phone number that rings a competitor. The engine guesses when the real answer is buried in a PDF or a stale listing, and the guess sounds confident.

In the field

A dental group had merged two offices a year earlier. ChatGPT still gave the old address and hours for the closed one, and Gemini listed a hygienist who'd left. A patient drove to a locked door. The fix started with logging exactly what each engine got wrong, with the right answer next to it.

Edge cases

• Answers that are dated, not wrong: old pricing, a since-closed location.
• A fact the engine credits to a competitor's page instead of yours.
• Safety-critical specs (pressure, compatibility, dosage-adjacent claims) where a confident wrong answer is the dangerous kind.

Why it matters

These get blurred together and they aren't the same. A business can get mentioned a lot and cited never, and that gap is the whole story. Measure against two specific rivals, not the market. A distributor scores against the two competitors who keep winning interchange queries. A roofer scores against the two firms ChatGPT keeps naming in its metros. A dental group scores against the two practices that own "near me" in its city. One number, tracked monthly.

In the field

An independent MRO supplier kept getting mentioned in "food-grade lubricant" answers but was never the linked source. Grainger's pages were. Citation share against Grainger sat near zero while mention rate looked healthy. That gap, not the comfortable mention number, was the thing worth fixing.

Edge cases

• A high mention rate hiding a citation share of zero.
• Scoring against the whole market instead of two named rivals, so the number never moves in a way you can read.
• Reading a single month as a trend. One run is a snapshot, not movement.

Why it matters

Reporting that you lost isn't useful. Saying why is. The usual causes repeat across verticals. The competitor's content is plain HTML the bots can read while yours is a PDF or a JavaScript app (a page that builds itself in the browser, which many AI crawlers don't run). Or there's an entity gap, where the engine isn't sure what you even sell or where you operate. For a multi-location roofer it's often that the city pages aren't reachable. For a dental group it's a name, address, and phone that don't match across listings, so the engine can't trust which office is which.

In the field

A roofing company's 60 city pages existed, but the only links to them sat in a JavaScript dropdown the crawler never ran. Competitors with plain linked city pages got cited in metro after metro. The one-line diagnosis, "our city pages are unreachable, theirs aren't," was the start of the fix list.

Edge cases

• A retrieval crawler (the bot that fetches a page to back up a cited answer) blocked at the firewall, so good content is never seen.
• An entity gap where the engine mixes you up with a similarly named business.
• Assuming it's a content problem when it's a reachability or trust problem.

Why it matters

The report is the wake-up call. The fix list is what earns its keep. Rank by what moves the business. A distributor fixes the interchange chart that loses real orders before the nice-to-have category page. A roofer fixes the city pages that cover its actual service area before the metros it doesn't serve. A dental group fixes the wrong hours on its busiest office first. The monitoring doesn't replace the fix. It justifies the fix, then measures whether it worked.

In the field

A regional electrical distributor's panel surfaced a dozen problems. We ranked them. The obsolete-breaker cross-reference (high-intent, high-revenue, fixable as one readable page) went to the top. A cosmetic schema gap on low-traffic pages went to the bottom. The list, not the raw report, is what the GEO team actually worked from.

Edge cases

• A long list with no order, so the cheap cosmetic fixes get done and the revenue ones don't.
• Fixes that need product-data, listings, or dev work the buyer can't resource yet. Flag the dependency.
• Treating every wrong answer as equal when one costs a sale and one costs nothing.

Why it matters

There's no standard definition of AI share of voice (your slice of the answers across the prompts you track), so you set one that's honest and stable, then hold it. The frame changes by business. For a distributor it's citation share on part-number and cross-reference queries against named rivals. For a multi-location roofer it's presence across the metros it actually serves, weighted by where the revenue is. For a dental group it's "near me" and treatment queries per office. The discipline is reporting the same thing the same way each month so a trend means something, and not claiming the number proves more than it does.

In the field

A roofing company with offices in eight metros wanted one "AI visibility" number. A flat average hid that it was strong in two cities and invisible in six. We defined the metric as metro-weighted citation share, so the report tracked presence where the trucks actually were. The monthly line finally matched reality.

Edge cases

• A single blended score that hides per-location or per-category collapse.
• Changing the definition mid-program so the trend line lies.
• Tying the number to revenue and claiming credit the data can't support.

Why it matters

Accuracy isn't a one-off audit. The engines drift and re-break. The stakes scale with the vertical. A distributor governs pressure ratings, compatibility, and interchanges, where a wrong answer ships the wrong part. A dental group governs hours, locations, and which services each office offers, because a wrong answer sends a patient to a locked door. A roofer governs service areas and licensing claims. You fix the source (the readable page, the Google Business Profile, the listing), then watch the panel to confirm the engine actually picked up the change.

In the field

A hydraulics distributor's monitoring kept catching an interchange stated backwards: the engine recommended swapping a higher-pressure part for a lower-rated one. Fixing the page once wasn't enough. The engine re-surfaced the old answer weeks later from a cached source. Governance meant re-checking until the correct fact held across all three engines.

Edge cases

• A correction that fixes one engine but not the others.
• An engine reverting to a cached or third-party source weeks after the fix.
• Wrong facts that come from off your own site (a directory, a reseller) where you can't edit the source directly.