The prospect reading your services page in San Francisco may have spent last quarter shipping a retrieval pipeline. Tell that person your firm scores sixty-one for AI visibility and they will ask what was sampled, how often, and against which model.
That reaction is the most useful thing to hold in mind here. In most markets an unexplained score buys a little credibility. In this one it costs some, because the audience knows what sits behind the number and how thin the ground under it is.
Your buyers built the machinery you are being sold a score for
An accountant in the Financial District, a recruiter working technology mandates, an executive coach with a practice full of founders — each sells to people whose profession is evaluating systems that produce text. They have written the prompt that pulls documents into a context window. They have watched a model cite a source saying the opposite of the answer. They have opinions about chunking.
That makes the constraint unusual. Elsewhere the risk is that nobody notices what you did; here it is that somebody notices exactly what you did and marks you down for it. A page assembled to appeal to a language model has a texture, and this reader knows it: front-loaded definitions, redundant restatement, question-shaped headings that answer nothing.
- The reader can smell the optimization. Text written to be retrieved rather than read is obvious to anyone who has built a retrieval system, and it is not a flattering recognition.
- Unverifiable claims get checked. A figure on your site with no stated method is an invitation to test it, and this audience accepts more often than any other.
- Trust is granted for limits, not confidence. Saying plainly what you do not know reads as competence; saying everything firmly reads as marketing.
- The owner is usually not technical. The dentist or the general contractor cannot audit a vendor's claim, yet is judged on the result by people who can.
What actually changes when the answer is generated
A results page is a list of documents in an order. The object is stable: your page occupies a slot or it does not, the slot can be observed, and two people searching from one place see roughly the same list.
A generated answer is a different object. It is written on the spot, from material the system fetched a moment earlier, in response to that exact phrasing. Ask twice with different wording and the answer differs, sometimes in which sources it leans on. There is no slot — there is a paragraph, and near it, sometimes, a few attributing links.
The unit of measurement moved
Position was a property of a page. Inclusion is a property of one answer, to one phrasing, at one moment — which makes any sampling scheme an approximation.
- Rephrasing changes what gets fetched
- Consecutive runs can differ
- Nothing corresponds to "we were third"
The answer often ends the session
A summary can satisfy the question without the reader visiting anybody. Exposure of that kind leaves no click behind, so your analytics record does not contain it.
- No entry in the click column
- The reader may return later by name
- Nothing joins those two events
Useful beats close
A system assembling an answer needs sentences it can lift and stand behind. A page thick with unsupported adjectives offers nothing liftable, however well it matched.
- Specific figures survive summarization
- Vague reassurance does not
- Exclusions are unusually citable
Credit is discretionary
A system can use your page to shape an answer and name someone else, or nobody. Use and citation are related events, and only one of them is ever visible to you.
- Influence without attribution is common
- Patterns differ between products
- Neither is published by anyone
Cited is not ranked, and the gap runs both ways
The instinctive assumption is that the answer layer draws from the top of the results page, so ranking well delivers citation automatically. That is partly true and misleading in both directions. Pages outside the first few positions turn up fairly often, usually because they state plainly what the higher pages talk around; pages ranking first get passed over when they lead with positioning instead of substance.
For a small practice that asymmetry is one of the few encouraging facts here. Competing for a slot against a well-funded firm is a resource contest you may lose. Being the only firm in the city publishing an actual hourly rate, an actual turnaround in days, or an actual list of what a fixed fee excludes is a willingness contest — and most of your competitors are unwilling.
| Property | Classic ranking | Citation in a generated answer |
|---|---|---|
| What is measured | Position of a document in a list | Whether a passage was used, and named |
| Stability | Reasonably steady day to day | Varies with phrasing and with the run |
| Official record | Impressions and clicks in your property | None published by anybody |
| What wins it | Relevance, authority, links, tenure | A statement worth lifting, and trust enough to lift it |
| Effect on the reader | A visit you can attribute | Frequently no visit at all |
| Who can verify | You, in the console | Only by asking and reading the answer |
Read that as a division of labor. The left column still pays the bills: measurable arrivals from people who chose you over what else was on screen. The right shapes a conversation you cannot observe. Treating it as a replacement is how a practice ends up with a handsome report and a quiet phone.
Which pages get pulled into an answer
Nobody outside those companies knows the retrieval criteria, which deserves stating flatly before any pattern is described. What follows is inference from observed behavior, not documentation, and it will age.
With that said, material appearing in generated answers about local professional services tends to share properties. It states something concrete, in a sentence that survives extraction, from a source with enough corroboration elsewhere to be worth trusting — and it answers the question asked rather than the one the business wished had been asked.
- Numbers with their conditions attached. "Between $2,400 and $4,000 for a standard filing, depending on entity count" is liftable. "Competitive pricing" is not.
- The answer early in the document. Not as a formatting trick but as a courtesy that happens to be machine-legible. If the answer is on the page, put it where a reader would look.
- Corroboration off your own domain. A firm named in a trade directory, a licensing register and two local publications is easier to trust than one existing only on its own site.
- Explicit limitations. "We do not handle international tax" is more citable than any positive claim on the page, and it prevents the wrong enquiry arriving.
How a visibility estimate is built, and what it is worth
Since there is no meter, an estimate has to be constructed. The approach is the same everywhere: assemble questions a customer might ask, put them to generative systems, record whether the domain appears, is named or is described, repeat on a schedule, and score the result.
Every step hides a decision that changes the output. Which questions, how many, in whose phrasing. Which systems, from where, with what account state. How many runs per question, given that one run samples a stochastic process. What separates a mention from a citation, and what weighting turns a pile of appearances into a number between zero and a hundred.
Where the model places you among rivals
A generated reading of position, peers and gaps — inference throughout, and presented as such.
- Competitiveness score with a market circle. Rival domains sorted into top tier, mid tier and niche, arranged around your position rather than listed flat.
- Model-written market context. How the domain is positioned, an estimate of its traffic, and where the model believes room exists.
- Competitor strengths and content gaps. What rivals cover that you do not — the one output here you can act on the same afternoon.
What the estimate is good for is comparison against itself and against neighbors. If you and four competitors are sampled the same way on one day, the ordering carries information even when the absolute numbers do not: directionally usable, precisely worthless.
Write things that can be checked, and let the citation follow
The suggestions falling out of all this are unglamorous, and their virtue is that each also improves the page for a human deciding whether to call you. None is a trick, and none stops working if the mechanism changes next year.
Where the model says to write next
Query research, intent classification and the pages flagged as leverage — a shortlist to argue with.
- Query research and intent classification. Phrasings grouped by what the person wanted, which is the fastest way to see which questions your site never answers.
- Leverage pages. Documents the model marks as worth expanding or linking from, usually because they already collect traffic on an adjacent subject.
- Human judgment still required. The model does not know you stopped taking a category of client in March.
The rate card you keep postponing
Ranges, with the variables that move them. The technical buyer wants to self-qualify before contacting you; a system assembling an answer wants a sentence with a figure in it. One page serves both.
- State what is included and excluded
- Give turnaround in days, not "quickly"
- Date it, and revise the same page
What you decline to do
An explicit scope boundary is among the most citable sentences a practice can write, and it removes the enquiries that cost an hour each to decline politely.
- Client sizes you are not set up for
- Specialties outside your remit
- Work you refer out, and to whom
The process, in steps with durations
What happens in week one, week two, week six. Concrete sequences extract cleanly, and they answer the question every prospect has and few ask out loud.
- Number the steps
- Attach a realistic duration to each
- Say what delays it
Anything written for the machine
Padded question-headings, definitions restated four ways, paragraphs existing to contain a phrase. This audience spots the pattern instantly and reads it as a signal about the firm.
- No synthetic FAQ nobody asked for
- No repeating one claim per section
- If it reads badly aloud, cut it
Folding it into a monthly report without overclaiming
The practical question is how a figure like this sits beside the ones that were counted. In its own section, labeled, with the method in one sentence, and never added to anything.
| Line in the report | Where it comes from | How firmly to state it |
|---|---|---|
| Clicks and impressions | Your verified property, two-day lag | As fact, within its fence |
| Average position by query | The same property, averaged | As fact, per named query only |
| Tracked rank for a term | A controlled sample of the results page | As a sample, with its conditions |
| Competitor overlap | Shared keywords and authority scores | As an external estimate |
| AI visibility reading | Sampled questions put to generative systems | As inference, with the method named |
| Market context narrative | Model-written from domain signals | As a hypothesis to test |
One further discipline. Whenever the inferred figure moves, look for a counterpart in the counted data before concluding anything. A rise in visibility with no matching movement in impressions on the same subject is a reason to distrust the reading, not to celebrate it. Where the two agree you have something; where they disagree, the counted one wins.
Both sides export together, which is the modest case for one account rather than three: the report builder with CSV and JSON export handles ten thousand rows, with a rendered PDF at two hundred and fifty for the version a client reads. Our written material covers the counted side, and intent classification across your own query set narrows the list rather than settling it.
Questions that arrive once somebody technical is in the room
Can we see how often an answer engine cited us last month?
No, and neither can anybody else. There is no log, no dashboard and no API from the answer engines reporting it. Every figure on the market comes from asking a set of questions and recording what came back — a survey with a small sample, and it should be described that way whenever it is shown to anyone.
Should we restructure our pages so a model parses them more easily?
Structure them so a person finds the answer, which produces most of the same effect: clear headings, a direct statement early, tables where you compare things. Avoid the second layer — sections existing only to be extracted. Prospects here recognize that pattern from their own work, and it costs more with them than it gains elsewhere.
Traffic fell but enquiries held steady. Is the answer layer eating our clicks?
One plausible explanation, and not the only one. Check whether the lost impressions were on informational phrasings rather than commercial ones, whether a competitor published something new, and whether a client category in your book has simply shrunk. Here the last of those is common enough to check first.
How long before a real pricing page shows up anywhere?
For ordinary search, four to eight weeks before movement means anything, assuming prompt discovery. For generated answers there is no reliable interval to quote, because there is no way to observe the start of it. Judge the page on whether prospects stop asking for a ballpark on the phone — that signal arrives sooner and is worth more.
A vendor showed us a competitor's visibility score. How much should that worry us?
Ask three things: how many prompts, how many runs each, and what counts as a mention. Vague answers mean the number is decoration. Specific answers, with the same procedure run against both of you on one day, make the ordering informative — still not enough to justify a budget by itself.
The part of this that will still be true in two years
Every mechanical detail here is provisional. Which systems dominate, how they attribute, what they reach for — none of it documented, all of it inferred, and it has already changed several times. Advice pitched at the current mechanism has a short life.
What does not expire is the underlying position. A firm whose site carries real prices, real timelines and real statements of what it declines to do is easier to summarize accurately, easier to trust and easier to choose. That was true when only people read your site, it is true of a system assembling an answer, and it will be true of whatever follows.
It is also the cheaper path. Out-publishing a firm with a content budget is a losing position. Being the one accountant in the neighborhood who publishes a fee schedule, or the one contractor who states what a kitchen costs and why the range is wide, is a position nobody can outspend you on — because money is not what blocks it. Nerve is.
To see the inferred readings sitting next to the counted ones rather than in separate documents, connect a verified property and open the panel. The generative research views, the analytics record and the rank tracking read from one account, with campaign automation from $149 per month per domain and a managed tier at $500 that includes people rather than only software; our engagement outlines cover the version where somebody else runs the monthly pass. Either way, start with the three sentences a prospect cannot find on your site today, and treat the visibility reading in the AI search landscape as the last thing you look at, not the first.