A generated answer does two things at once. It states a position, and it says whose position it is. For an organization whose entire standing rests on being the source other people cite, the second of those is the one worth losing sleep over.
Something now sits on top of the ranked list. Type a question and a paragraph is composed for you, with a narrow band of sources underneath. Most commentary concerns the missing click. In this city the more consequential change is that a machine has begun making attributions, and attribution is what a great many Washington organizations sell.
Classic ranking has not gone anywhere. Impressions, clicks and position history remain the only figures here with real counting behind them. Above them a second layer has formed, and every instrument aimed at it — the generative research screens inside the rebuilt Semalt workspace included — produces inference, not measurement.
The answer arrives already written, and it names somebody
A results page never asserted anything. It set out candidates and left the assessment to the reader. That is why click-through rate meant something: a person chose, and choices can be tallied. Move the choosing inside the system and the tally goes with it.
The bigger difference is that a written answer takes a side. Ten links about a proposed rule cannot be wrong. A paragraph summarizing it can be, and it arrives with a source name attached, which turns a summary into a quotation. For any organization whose analysis is its output, that is not a cosmetic change to a marketing channel.
- Questions arrive fully formed. People write out the whole situation when a machine is reading — the statute, the year, the exception — so a two-word term becomes a sentence with three conditions in it.
- Whoever still clicks is verifying. They already have an answer and came to see whether your page supports it — nothing like a browsing visitor.
- You are quoted without being told. No notification exists and no log is kept, whether or not the summary resembles what you published.
A place in the queue, and a line in the footnote
The intuition is that these amount to the same thing: rank highly and you get quoted. The relationship is real but loose enough that the two must be tracked apart, which is why a second family of screens exists.
Position is a queue. One organization holds each slot, and every gain is somebody's loss. Being drawn on is not arranged that way: one answer can assemble material from four places, each supplying a piece. What earns a document its piece is a passage settling one narrow question cleanly enough to lift whole. A page sitting eleventh with a dated, conditioned, plain sentence gets used ahead of a second-place page whose nearest equivalent paragraph is about commitment and mission.
| Property | Position in the list | Use in an answer |
|---|---|---|
| Unit being judged | A whole page against a term | One passage against one sub-question |
| How many can win | One organization per slot | Several, in the same paragraph |
| What decides it | Relevance, links, technical health | Whether a sentence can be lifted and stood behind |
| What failure looks like | You drop three places | Somebody else is credited |
Granularity is the second split. Rank tracking follows one term at a time; being drawn on happens one statement at a time, so a well-built explainer can feed thirty differently phrased questions while ranking notably for none. The planning input is a list of questions, not keywords.
In this town, attribution is the product
Take stock of who publishes here. Trade associations that exist to be the authoritative voice of a sector. Research institutions whose funding rests on being quoted in committee. Standards bodies whose documents are the reference. Professional societies, advocacy groups, law firms whose client development runs on published analysis, security vendors whose research desks issue findings the field repeats.
None of them is chasing traffic in any ordinary sense. What they accumulate is standing — being the place a serious person goes for the number. Correct citation is not vanity for them; it is the mechanism by which the enterprise works, and it has been mediated for a century by editors and conventions that assign credit deliberately.
A generated summary assigns credit too, by a procedure nobody outside the vendor can inspect and nobody at all can appeal. That is the whole story in one sentence, and it is why this belongs on the agenda at organizations that would otherwise leave search engines to a junior staff member.
Given to somebody else, or handed back unrecognizable
Two things go wrong here, worth separating because the responses differ. In the first, the substance is yours and the name is not. In the second, the name is right and the substance has been compressed until it says something you would not have signed.
Credit lands on a neighbor
Your survey, your methodology, your finding — attributed to whichever organization restated it most quotably.
- Coverage of your work outranks the work
- Your own page never states the finding plainly
The position loses its conditions
You wrote that a cost rises under one scenario. The summary says the cost rises, full stop, and puts your name on it.
- Scenario, baseline and date stripped away
- A conditional recommendation read as endorsement
The second failure is the more dangerous, because it is invisible from inside. Your site did not change, and your report still holds every qualification you placed in it. What circulates is a compression that drops them, and you may first hear of it from a member asking why you now hold a view you do not.
| What a summary does | Who it exposes | The move that actually helps |
|---|---|---|
| Drops the scenario a projection depends on | Research institutes, policy shops | Put the scenario inside the same sentence as the number |
| Summarizes a legal reading with no jurisdiction | Law firms, lobbying and compliance advisers | Name the jurisdiction and the date in the heading |
| Turns a scoped finding into a general claim | Security research desks, government-technology vendors | Publish the scope, the sample and the test window |
Read down the third column and notice what never appears there. Nobody is writing to a model vendor, because no dependable channel of that sort has been built. Each remedy is the same act carried out on material you already possess: make the accurate, fully conditioned wording the easiest thing to lift. The answer is beyond editing. Its raw material is not.
Which documents an answer engine picks up on a policy question
The familiar complaint about answer engines is that they favor directories, roundups and review sites. On consumer questions that is broadly accurate, and a hotel or a trades company in Arlington will meet it exactly. On the questions this city's institutions care about, the composition looks rather more favorable.
Primary and official material
Statute and regulation as published, agency guidance, registries, dockets, and the official record of proceedings.
- No placement can be bought here
- Where your filings and testimony already sit
Named analysis with a date on it
Institute publications, association guidance, firm memoranda and society standards — provided they are signed, dated and specific.
- Realistically inside your control
- Loses instantly if undated or unattributed
The aggregator layer thins out here for a structural reason: no directory holds the comment deadline, the eligibility threshold or the effective date. Nobody has built a business out of restating those accurately, which leaves the gap to whoever publishes them plainly.
The jurisdictional problem from ordinary local search follows you up into this layer and worsens. Ask an assistant what a business must file before trading in the Washington area and you may get a confident, coherent answer describing exactly one of the District, Virginia or Maryland, with no signal that the other two differ. A page naming its jurisdiction in the heading has a chance of being quoted correctly. One written for "the DMV" as a single regulatory space invites the flattening, then carries the blame for it.
How a visibility estimate is actually assembled
Because no citation data is published, any number describing this layer is built from outside. Understanding how is the fastest way to calibrate what it can be asked to do.
Where the estimate comes together
Generative market research: how a model appears to see a domain, rather than where a crawler places it.
- A competitiveness score with its comparison group shown. A Market Circle sorts rivals into top-tier, mid-tier and niche bands, so the peer group is visible rather than assumed.
- Query research and intent classification. What people in your field are asking and what they appear to want, grouped by purpose instead of volume.
- Leverage pages and missing ground. Documents the model marks as worth expanding or linking internally, plus subjects your rivals cover and you have never written about.
- One rolled-up visibility value. A single figure for standing in the generated-answer landscape — the most quotable output here, and the one needing most care.
The procedure behind that last figure is neither secret nor complicated. Questions are put to models. Whatever comes back, answer text and citation strip alike, is scanned for domain names. How often and how visibly each turns up becomes a weight, the weights are ranged against a named peer group, and the result is expressed as one number. Every step is defensible. None of it watches a real person read anything.
What the estimate can carry, and the rooms it must not enter
None of that makes the figure worthless. Keep the questions and the peer group frozen, take three or four readings before concluding anything, and it will tell you something genuine about which way things are moving. What it cannot survive is being lifted into a document that expects provenance — which matters more here than almost anywhere, since these organizations spend their lives filing into exactly those settings.
That final tile needs a caveat. Four to eight weeks describes when link and ranking work typically starts to register. It says nothing about appearing in a generated answer, which cannot be scheduled, since model updates arrive unannounced.
Write so that the paraphrase is still your position
The content work is unglamorous and mostly concerns documents already on your own server. A model wants a passage it can extract without interpreting. A visiting analyst wants a page that survives checking. Conveniently, the same writing satisfies both.
- Keep the condition and the claim in one sentence. A number carrying its scenario, baseline and date survives being moved. Park the caveats two paragraphs down and only the bare number travels — under your name.
- State your finding before anyone restates it. If the clearest sentence describing your research sits on somebody else's site, expect the credit to follow it there.
- Put a byline and a date on it. An author who holds an actual job title, the day it went up, a line noting any revision. Their absence is the first thing a careful reader registers.
- Name the jurisdiction in the heading. A document saying which of the three regimes it covers is quotable; one that does not gets flattened into whichever the model assumed.
- Publish the numbers only you hold. Thresholds, deadlines, fee schedules, membership criteria, survey method — your best paragraphs, because no aggregator can manufacture them.
Reporting is where this stays honest or quietly stops being so, and the failure mode is predictable enough to build against. The estimate lands beside clicks, shifts nine points, and by the second quarter somebody defends that shift in a meeting as though a machine had counted it. Good intentions will not prevent that. Fixing the position of the number, and the wording that always accompanies it, will.
Counted first, estimated second, labeled every time
An order of presentation that admits the new layer without letting it crowd out figures that were genuinely measured.
- Hard figures open the document. Clicks, impressions, terms entering or leaving the top three, ten and thirty — counted, each with a window attached.
- The estimate follows, wearing its method. Date of reading, question set, peer list, one sentence of procedure — every month, in the body, not in a footnote written once and forgotten.
- Close on what was done, not on adjectives. Four pages re-dated and attributed, two findings restated plainly on our site, one jurisdiction given its own document.
The mechanics are handled by a report builder you configure and brand; the judgment stays with you. Both layers feed on the same material, which is the argument for housing them together: dynamics names the terms crossing into the top ten, the rival screens name who else holds that ground, and the query work shows how the same need gets typed when a machine is expected to read it. Our service overview describes how the review runs on a live property.
Questions that come up the first time this is raised
Can we find out how often an assistant has cited us?
No, and neither can any supplier. Sampling is the only route: ask the same catalogue of questions over and over, then record which domains surface in the reply and in the strip beneath it. What you get is a position against named peers, never a count. Anyone displaying a citation total should be asked what they divided by.
An answer credited our research to another organization. Is there anything to be done?
Not through an appeal; no such channel exists. Look instead at which page states the finding most clearly. If a secondary write-up says it more plainly than your report does, the credit is following the clearer sentence. Publish a short, dated, plain statement of the finding on your own site and give it time.
Our funder wants a figure for AI visibility in the annual report. What do we give them?
A description rather than a score. Say what you sampled, when, against which peers, and which way it moved, in wording that makes plain the reading was constructed. If a number is unavoidable, print the method beside it. A grant report carries an audit trail, and an unlabeled inferred figure is a liability inside one.
Why did our estimate move when we published nothing?
Most likely because something changed on the other side. Model versions get replaced, source preferences shift, and one question asked twice can return different material. A single reading settles nothing, and the question set must be identical between comparisons.
Set the novelty aside and what remains under your hand is short. The precision of your own sentences. Whether anyone signed and dated them. Whether the facts nobody else possesses have been written down at all, and whether each document announces the jurisdiction it applies to. Modest levers, but real ones, and identical to the ones that serve a reader who opens the footnote.
Everything else belongs to somebody else: this quarter's preference among source types, whether an official register beats your commentary on a procedural question (it generally does, and generally ought to), what next month's update rearranges. To see how your domain sits against the peers you actually compete with, connect the site and open the generative research views, then read the blog for the analytics and indexing companions to this piece. Look at the citation strips before the score. In a town that argues about authority for a living, the finding worth having is rarely numeric. It is discovering which document is currently talking on your behalf, and whether you would sign your name to it.