Direct answer: That’s normal — and it’s the single biggest reason one-shot self-checks mislead. Shoreline Digital Agency’s August 2026 newsletter — built on 31,495 AI prompts across ChatGPT, Gemini, Perplexity, and Claude — points to a citation study finding a brand’s AI visibility can vary by more than 4.5 times across closely related topics in the same niche. “Best plumber in Katy for a slab leak” and “plumber Katy” are, to an AI engine, two completely different questions that retrieve completely different evidence. If you checked once, with one phrasing, you haven’t checked.
The owner who concluded too fast
Related: Your Website Doesn’t Need More Traffic. It Needs to Be the Answer.
Here’s the scene: an owner opens ChatGPT, types “plumber near me” or “best plumber in Katy,” doesn’t see his company, and concludes “ChatGPT doesn’t know me.” He might then spend money fixing a problem he doesn’t have — or ignore a problem he does.
CI Web Group’s checklist contains the corrective in one line: “‘Best plumber in Katy for a slab leak’ beats ‘plumber Katy.’” The specific, question-shaped phrasing — service, city, job type — retrieves different sources, triggers different reasoning, and names different businesses than the keyword-shaped phrasing. The homeowner asks questions. The owner tested keywords. They weren’t running the same query. (CI Web Group)
Shoreline Digital Agency’s August 2026 newsletter, “31,495 AI Prompts Later: What Actually Gets a Local Business Cited,” points at a citation study that found a brand’s AI visibility can vary by more than 4.5 times across closely related topics in the same niche. A business cited constantly for one phrasing can be invisible for a near-identical one. Shoreline’s own prescription is structural: “Coverage across many phrasings, not one keyword. Businesses that show up for ‘who does X in Y,’ ‘list of X near Y,’ and ‘recommend an X in Y’ all at once get cited far more often than ones chasing a single exact-match term.”
More than 4.5x. Same business, same niche — over four-and-a-half times the visibility depending on how the question is worded.
Why phrasing changes the answer
An AI answer isn’t a ranking; it’s an assembly. The engine takes the question, searches the web for sources matching that specific question, and builds an answer from what comes back. Change the question and you change the retrieval:
- “plumber Katy” retrieves directory-shaped results — list pages, aggregators, whoever dominates the category pages.
- “Best plumber in Katy for a slab leak” retrieves answer-shaped results — Reddit threads where someone asked exactly that, review language mentioning slab leaks, service pages that actually discuss slab leak work.
- “Who should I call for a slab leak under my foundation in Katy?” retrieves yet another set — possibly forums, possibly Q&A content, possibly nobody, if nothing on the web answers it that specifically.
Each phrasing is a different retrieval against a different slice of the web. Your business can be the best-evidenced answer for phrasing two and absent from phrasing one, because the evidence the engine found differs. This is also why the engines differ from each other on the same question — different retrieval, different wells.
Note what this is not: it’s not geography (that’s the “new towns I’ve never ranked in” phenomenon — a different question). It’s not randomness, exactly — though non-determinism means the same phrasing can vary run to run. It’s phrasing variance: systematic, repeatable, and measurable.
The matrix: test like a homeowner, not like an SEO
Run this matrix on the cadence in LLM visibility measurement.
Stop running one check. Run a matrix. For each service × city you care about, test 6–12 phrasings across the question shapes homeowners actually use:
- Direct: “Who’s the best [trade] in [city] for [job]?”
- List: “List [trade] companies near [city].”
- Recommendation: “Recommend a [trade] in [city].”
- Problem-first: “My [specific problem] in [city] — who should I call?”
- Comparison: “Which [trade] in [city] is best for [job]?”
- Urgent: “Emergency [trade] [city] right now.”
Run each in ChatGPT, Perplexity, and Gemini (or Google with AI Overviews). Screenshot everything with the date. Record three statuses separately: named, cited-but-not-named, absent. Do it monthly.
What you’ll find: a pattern, not a verdict. Maybe you’re named for problem-first phrasings but absent from list phrasings — that tells you your review language is strong but your directory presence is thin. Maybe you’re cited everywhere but never named — that tells you the engine uses your content but trusts someone else’s entity. The matrix turns “am I visible?” into “visible where, for which questions” — and each cell suggests a different fix.
What to build: one page per job, in the customer’s words
The fix for phrasing variance is coverage, and coverage comes from pages. A business owner quoted in a July 2026 Reddit thread (via pluspoint.io’s write-up) described the move exactly: “I made a page for each of the top five types of jobs… I’ve been getting a lot of people who found me on AI.” The move is five pages, not one services page listing five jobs. Each page answers one job’s questions in quotable sentences: what it costs in your area, how long it takes, what happens on the first visit, what goes wrong.
Outside research the newsletter lines up backs the shape: answer-first content earns meaningfully higher citation rates than generic copy, and pages that open with a clear, direct answer — the “answer capsule” — outperform pages that make the AI assemble an answer from scattered text. Structured formats (Q&A, clear headings) beat dense paragraphs. A large-scale analysis of tens of thousands of brands found brand mentions across the web correlate with AI citation about 3x more strongly than backlinks. A benchmark found a complete business profile makes you roughly 70% more likely to appear in AI local recommendations. And separate local ranking research puts review signals at roughly 20% of local ranking weight now, up from 16%.
None of that is phrasing-specific. All of it is coverage: be the best-evidenced answer for as many phrasings as possible, instead of optimizing for one.
The discipline: never decide on one screenshot
This is the behavioral point, and it’s the one that costs owners money. The failure modes:
- One check, wrong conclusion: “ChatGPT doesn’t know me” (tested one keyword phrasing, once).
- One check, false confidence: “We’re all over ChatGPT” (asked the exact question your homepage answers, once).
- One check, wasted spend: buying a “visibility package” because a vendor’s single screenshot showed you absent.
That 4.5x finding is the antidote. Any single observation sits somewhere on a range wider than 4.5x, and you don’t know where. The matrix is the minimum viable measurement. Monthly is the minimum viable cadence — answers shift as sources get re-crawled and engines retune retrieval (August 2026’s ChatGPT retrieval change moved citation shares overnight).
Pair this with the interrogation habit: when the matrix shows you’re absent for a phrasing where a competitor is named, ask the engine “Why did you recommend them?” for that phrasing. The variance tells you where to look; the interrogation tells you what to fix.
The line to remember
Your visibility isn’t a fact. It’s a range — more than 4.5x wide — across the phrasings your customers actually use. One check proves nothing. Run the matrix monthly, build one quotable page per job, and stop letting a single screenshot make decisions for you.
FAQ
Q: How many phrasings do I really need to test? A: Six to twelve per service × city, across the question shapes above, in three engines. That’s 20–40 checks monthly — about half an hour once you have the template. Start with your top three services and top three cities.
Q: Should I phrase my website to match? A: Write pages in the customer’s question language, not keyword language — but write for the customer, not the matrix. A page titled “Slab Leak Repair in Katy: Cost, Process, and What Happens First” naturally covers five phrasings. Gaming phrasings is a treadmill; answering questions is durable.
Q: Does this variance mean AI answers are random? A: No. Phrasing variance is systematic — the same phrasing retrieves similar evidence repeatedly. Same-phrasing re-runs wobble too, so don’t treat any single check as final. That’s why the matrix works: patterns emerge across runs even though individual answers wobble.
Q: My competitor shows up for every phrasing. How? A: Coverage: they’re quotable across more question shapes — more specific reviews, more directory completeness, more answer-shaped pages, more mentions. Interrogate the engine (“why did you recommend them?”) on the phrasings where they win and you’ll see which evidence you’re missing.