Everybody here has read the same four posts about optimizing for language models, and a good number will ask you for a visibility number by Friday. That number does not exist in the form they want, and the tactic they hope to hear about is the one thing that reliably fails.
A second layer has settled over search. Ask a question and something writes a paragraph, with a thin row of sources underneath. The paragraph is the product; the sources are what concerns you.
Ordinary ranking has not gone anywhere, and its figures remain the only ones with anything actually counted underneath. Above them sits a layer nobody tallies, described by instruments — the generative research views in the rebuilt panel among them — whose output is inference.
The answer arrives assembled, and nothing about it is countable
A results page never made an argument. It set out candidates and left the deciding to the reader, and deciding produced a countable event — a choice made on infrastructure Google owns.
An assembled answer removes the choosing. What replaces the click is inclusion: whether a sentence of yours ended up in the paragraph, and whether your domain appears in the strip beneath. Nobody logs that for you.
- The unit of success changes size. A ranking is won by a page against a phrase. Inclusion is won by a passage against one part of a question.
- Several sources win at once. One answer routinely draws on four or five documents, each supplying a fragment. Nothing here resembles a ranked queue.
- Questions arrive as full scenarios. People type whole sentences at a machine, so a two-word term shows up carrying a budget and a constraint.
- Surviving traffic differs in kind. Whoever clicks after reading an answer is checking a claim, not browsing. Fewer visits, arriving further along.
Being cited and being ranked are not the same competition
The instinct is to treat these as one thing: rank well, get quoted. A relationship exists, and it is loose enough to justify a separate layer of analysis. Pages sitting in unremarkable positions get quoted constantly, and pages in the top three get passed over.
What decides it is whether a passage can be lifted out and stood behind — a sentence resolving one part of the question cleanly, carrying its own conditions, not needing the page around it. A ninth-placed page with a dated figure and a named constraint supplies that. A second-placed page whose matching paragraph praises its own commitment to partnership does not.
| Property | Position in a ranked list | Use inside an assembled answer |
|---|---|---|
| What competes | A page, against a query | A passage, against a fragment of the question |
| How many win at once | Exactly one per place | Four or five documents together |
| What tips it | Relevance, links, technical health | Whether a statement survives extraction |
| Status of the figure | Counted and reported to you | Inferred from sampling you commissioned |
Scale is the other difference. Ranking is monitored phrase by phrase; quotation happens claim by claim. A planning document built as a column of keywords is therefore describing a contest you are not in.
The instinct to game this is the reason it will not work for you
This is where the local situation rearranges everything above. In most markets an article like this reaches an owner who has never thought about search mechanics. Here it reaches a room where a third of the people once ran growth for a product company and everyone skimmed a thread on prompt-level optimization.
That produces a specific request: a strategy for appearing in generated answers, framed as a lever, with a number attached so it tracks beside paid acquisition. Intelligent, and aimed at the wrong target. Both conditions that make a lever worth pulling are absent — the mechanism is undisclosed, and the measurement is not measurement.
The tactical reading
Treat inclusion as a channel with knobs. Find the pattern, apply it at volume, report the climb.
- Assumes a stable, discoverable selection rule
- Assumes the score has a real denominator
- Rewards whoever publishes fastest, briefly
The structural reading
Treat inclusion as a consequence of holding facts nobody else publishes, in quotable form.
- Survives model updates nobody announced
- Works on the questions with money behind them
- Compounds instead of decaying
The pattern-chasing version has a further problem here. Whatever trick is circulating this quarter, your competitors read the same thread on the same day. A shared tactic in a technically literate category has a half-life of months, after which everyone has done it and nobody is differentiated. Easy moves are consumed on contact here.
A second local problem sits underneath. Much of the demand comes from firms and people who arrived recently, and the words they use for your category move faster than in older markets. A question set built eighteen months ago partly describes vocabulary that has moved on, so a score computed against it measures questions fewer people now ask.
Where the answer's raw material comes from in your category
Composition of the source strip varies enormously by question type, and that variation is more useful than any aggregate score. Consumer-shaped questions pull directories and roundups. Technical questions pull documentation and whoever wrote the thing down.
| Question type | Typical sources | Realistic role for your site |
|---|---|---|
| Who should I hire for this in Austin | Directories, review platforms, listicles | Weak — a property of the question |
| How does it work, and when does it fail | Documentation, vendor guides, technical writing | Strong, if the page states conditions |
| What does it cost and what drives the range | Whoever published real numbers | Primary — almost nobody publishes this |
| What does the state or the city require | Agency pages, trade bodies, published legal analysis | Strong, when dated and tied to a jurisdiction |
| Which vendor suits this constraint | Comparison content, forums, product docs | Moderate — depends on specificity |
Treat the table as triage. Two rows repay effort, because the answer sits in your own records and nowhere else. The top row repays none, and chasing it is where most of this budget disappears.
The middle rows are where a Central Texas business is unusually well placed, because the facts belong to whoever does the work.
Numbers from the job
Review timelines across three counties, what a trailer buildout costs this year, current filing turnaround at an agency.
- No directory carries it
- No budget manufactures it
- Needs a date attached
Claims about quality
Experience, responsiveness, partnership, commitment. True, unquotable, identical across the category.
- Nothing there to extract
- Every rival says it too
- Wastes the page it sits on
What a visibility figure is built from, and what survives the construction
Since no citation data is published, any number must be constructed, and knowing how is the fastest route to knowing what it can bear.
Where the figure is put together
Generative market research — what a model says about a domain, as against what a crawler decided about it.
- A competitiveness score with a stated comparison group. The Market Circle files rivals into top-tier, mid-tier and niche, so you can see who the score holds you up against.
- Query research with intent classification. The questions circulating in your field, sorted by what the asker wants rather than how many ask it. Everything downstream builds on this list.
- One rolled-up visibility value. A single figure for standing across the generative search landscape, and the output needing most care when quoted.
Nothing about the procedure is secret, and it should be spelled out wherever the figure appears. Fixed questions go to models on a repeating schedule; whatever comes back, prose and source row alike, is scanned for domain names. An appearance counts for more when it recurs and more again when it sits near the front. Those tallies are set beside a named rival list and compressed into one value. No stage of it records anything a customer saw.
That does not make it useless. Held to the same questions and competitors month after month, the series says something real about which way things are moving. The information lives in the slope, never in one reading.
One caveat on that last tile. Four to eight weeks is when ranking and link work usually starts to register. There is no equivalent window for being quoted, because that depends partly on releases nobody tells you about.
Be the source worth citing on something narrow enough to own
Strip away the novelty and the content work is old-fashioned and mostly about pages you already have. A model needs a passage it can lift without breaking it; a suspicious reader needs a page that survives the visit.
- Keep the caveat in the same clause. Write the range, the conditions and the year alongside the figure. Set them three paragraphs lower and only the bare number gets lifted — which is then attributed to you.
- Publish the figures only you hold. Lead times, permit realities, tolerances, seasonal capacity, what drives a quote up or down. No competitor copies what they never measured.
- Say where your approach stops working. Spelling out when you are the wrong choice reads as competence, and it produces the kind of bounded claim a model can reuse without risk.
- Attach a person and a calendar. Who wrote it, what they actually do, when it went up, when it was last touched. Costs nothing, and a technical reader looks for those four things first.
- Keep the vocabulary current. Where category names shift every couple of years, a page using the last generation's word is invisible to a question phrased in this one's.
The narrowness is the point, and the part this audience resists because it feels too small. Nobody becomes a credible source on an industry in general; better-funded organizations hold that ground. Becoming the source on one specific, checkable, repeatedly asked thing takes a quarter, and it is what gets extracted.
Putting this in a monthly report without pretending it was counted
The failure mode is predictable, which makes it easy to design against. The estimate appears in a deck beside counted figures, shifts by eleven points, and two quarters later somebody defends that shift in a meeting as though it had been measured. The layout did that.
Counted first, inferred second, labeled every time
A layout rule that lets the new layer into the report without letting it crowd out anything that was actually counted.
- Start with what was tallied. Clicks, impressions, positions, and which terms crossed into or out of each ranking band. Every one of those has a chain of custody.
- The estimate follows with its apparatus. Date, question set, rival list and a sentence of method — printed every month, not buried in an appendix.
- Hold the question set fixed, or say you changed it. Comparing two different question sets measures the question sets. Updating for vocabulary drift is legitimate and resets the series.
- Close on actions, not adjectives. Four pages dated and signed, one cost page written with real numbers, two listings corrected. Checkable next month.
The mechanics themselves are dull. The configurable report builder handles layout, logo and colors. Both halves feed off overlapping material — the dynamics screen listing which terms entered or left each band, and the rival-domain views showing who else occupies your terms. Our measurement notes take the counted half further.
Questions that come up
Is there a count of how often assistants quote us?
There is not, for you or for anybody selling you software. Tools ask the same questions repeatedly and record which domains turn up. The output is a position relative to a rival list you chose, never a running total. When a dashboard shows a hard number here, ask what it was divided by.
Is there a way to optimize prompts so we get included more often?
Not durably. The selection mechanism is undisclosed and changes without notice, so a formatting trick bets on a system nobody outside the vendor can inspect. Where every competitor reads the same material, the trick is neutralized within two quarters. Publishing checkable facts nobody else has does not stop working when a model updates.
Our score jumped nine points this month. What did we do right?
Possibly nothing. Check first that both readings used the same questions, competitors and number of runs; if any of the three moved, you are comparing two experiments. If they matched, you have one data point. Wait for the next before telling anybody a story about it.
An answer mentioned us, but cited a thin roundup page. Win or not?
Half a win, and worth doing something about. The mention counts, but it rests on a page you can neither edit nor vouch for. Chasing the roundup is the wrong move; publishing a signed, dated page of your own stating those facts with their conditions is the right one. What exists to be quoted is among the few parts of this you can change.
How often should the question set be rebuilt?
More often than in most markets. Where many buyers arrived recently and category names keep splitting, a set built eighteen months ago partly describes questions nobody asks now. Review it quarterly against your own query breakdown, which is observed rather than modeled.
What is actually in reach, and what is not
Put the novelty aside and the controllable part is short. How exact your pages are willing to be. Whether anybody signed and dated them. Whether the material only your firm possesses was ever written down in public. Whether listings elsewhere describe you correctly. Most companies have touched none of it.
The list beyond reach is just as short. This quarter's preference for one kind of source over another. Whether a directory beats you as raw material for a hiring question, which it generally will. Whether an unannounced release rearranges everything in October. Nobody can promise you a place inside a generated answer, because it is not theirs to hand out — and here, whoever is promised it will usually know that.
So the better question is not what our number is. It is this: on which three specific things could this company be the best available source in Central Texas within two quarters, and what would we publish to make that true? That has a plan attached. The score, at best, says afterwards whether the plan works.
To point the sampling at your own domain and the competitors that actually matter, open the dashboard and connect the property. Start with what is linked beneath the answers, not the score. What the exercise produces is almost never a number — it is finding which page currently answers for your category, and deciding whether you would rather it were yours. Where that becomes an editorial program, our engagement outlines cover the scoping, and the query research and intent views are where the question set starts.
