AI visibility research and tracking

The prompt set decides everything downstream. Track the wrong questions and every number after them is precise, comparable, and about nothing anybody cares about.

The failure has a recognizable shape: a set built from what the company wishes people asked. Why is Horizon Legal the best Austin firm embeds its own conclusion, so the answer tells you nothing except whether an engine will repeat a premise you supplied.

What should I compare before hiring an Austin injury lawyer is the same territory and a completely different instrument. It reveals the decision criteria people actually use, and it shows you which competitors get named when nobody has tipped the scale.

Track the questions people ask, not the ones you would like answered. A prompt containing your conclusion cannot test anything.

Cover the decision, not the category

Group the set into cohorts by decision stage rather than by topic. A cohort is the unit you will compare over time, so it has to correspond to something you would act on separately.

Take the language from people, not from a tool

The words that belong in prompts come from recordings of sales calls, support tickets, community threads, and the queries in module 2's research. All of these are records of somebody trying to solve the problem in their own words, which is what a prompt is.

Remove anything written to flatter. Prompts that only exist to produce a favorable answer waste the run and corrupt the set. If a question would embarrass you to show a customer, it is not research, and the number it produces will be quoted internally as though it were.

Keep the set small enough to run repeatedly. Five approved prompts run weekly produce a comparable series; fifty run once produce a snapshot nobody can compare against anything. Repetition is worth more than coverage here, which is the reverse of module 2's keyword work.

One more filter is worth applying before a prompt is approved. Ask what you would do differently depending on the answer. A prompt whose result changes nothing is costing a run every week to produce a number that decorates a report, and a set assembled without that filter fills up with them faster than anybody expects.

Build the prompt set

Five steps. The output is a small approved set grouped into cohorts, not a long list.

1. Collect the language people actually use.Sales calls, support tickets, community threads, and the qualified queries from module 2. Written language from real situations rather than category terms.
2. Group questions by decision stage.Discovery through post-purchase. The gaps in your coverage become obvious the moment the stages are written across the top.
3. Add the constraints that change the answer.Location, budget, eligibility, timing. These produce the specific answers worth diagnosing, and generic prompts produce generic responses about your whole category.
4. Remove prompts that assume their answer.Anything containing your conclusion or your brand as the premise. They cannot fail, which means they cannot inform.
5. Approve a small set and assign it to cohorts.Small enough to run on a schedule for months. The cohort is what you will compare, so it should map to a decision you would make separately.

Module 10 established that one answer is an observation under conditions. This lesson is about writing those conditions down before you start, so that two runs six weeks apart are measuring the same thing.

Without a protocol, every difference between runs has at least three explanations and no way to choose between them. The engine changed, the conditions changed, or the world changed, and a report that cannot separate those is describing its own inconsistency.

Write the conditions before the first run. A difference you cannot attribute to a controlled variable is not a finding.

Record the conditions, then the result

Engine, product mode, account state, location, date, and the exact approved prompt text. These are cheap to record at run time and impossible to reconstruct afterward, which is why they get skipped and why skipping them is expensive.

Account state is the one people forget. A signed-in account with history produces different answers from a clean session, and a team that runs tests from whichever browser is open is mixing two populations without knowing it. Decide which one you are measuring and hold it.

Repetition is what makes it evidence

These systems return different answers to identical prompts. Running each prompt once and comparing weeks produces a chart of variance with a narrative attached to it.

Horizon runs five approved prompts weekly in the same products and country, and the tracker holds ninety snapshots across them. That volume is what turns a mention into a mention rate, and a rate is the smallest unit here that supports any claim at all.

Set the cadence to the decision, not to the dashboard. Weekly is right for an active program. Monthly is fine for a stable one. Daily produces cost and noise in equal measure, and nobody acts on a daily AI visibility number.

Version breaks are findings, not gaps

Products change. A search mode is added, a model is upgraded, a default changes, and the answers shift for reasons that have nothing to do with your site.

When that happens, record a protocol version break at that date and treat the series as two segments. The alternative is what usually happens: the shift lands in the middle of a campaign, gets reported as campaign impact, and the mistake is undiscoverable a quarter later because nobody wrote down what changed.

Define the classification rules in advance

Decide before the first run what counts as a mention, a recommendation, a citation, and a competitor appearance. Written afterward, these definitions bend toward whatever the data happened to show.

The edge cases are where the definition earns its keep. Does a mention of your parent company count? Does a citation of a third-party page describing you count as your citation? There is no universally right answer, and any consistent answer is better than deciding case by case for eight months.

Preserve every raw response and its cited sources alongside the classification. The classification is a judgment somebody made; the raw text is the evidence, and only the second one survives a disagreement.

Write the protocol

Six decisions recorded once, then held. Changing any of them starts a new series rather than continuing the old one.

1. Fix the engines and product modes.Named specifically, because a product's search mode and its default mode are different instruments returning different answers.
2. Fix the account state and location.Signed in or clean, and which market. Both change answers materially and neither is visible in the output.
3. Approve the prompt text and lock it.Character for character. An edited prompt is a new prompt, and comparing it to its predecessor is comparing two questions.
4. Set repetition and cadence.Enough runs per prompt to produce a rate rather than an anecdote, on a schedule matched to how fast you would act.
5. Write the classification rules, including edge cases.Parent companies, third-party pages, partial matches. Decide once, in advance, and apply it consistently even where it is unflattering.
6. Preserve raw responses and sources.Every run, verbatim. The saved text is what lets somebody re-derive your numbers, and re-derivable numbers are the only kind worth reporting.

The diagram above separates five reasons a brand is absent from an answer. This lesson is about the distinction underneath them, which decides where the work goes: whether you have a source gap or an answer gap.

An answer gap is about the response. You are missing, misdescribed, or framed badly. A source gap is about the evidence the engine had available: nothing useful existed, or what existed contradicted itself, or somebody else's page was simply better material.

A source gap is fixed by evidence existing somewhere. An answer gap may not be yours to fix at all.

Read the citations before the answer

The cited sources tell you more than the text does. Horizon's tracker shows three cited sites across the latest runs, of which one is theirs and two are other sites. That single line is the diagnosis: the engine is finding material on this topic, and mostly not from them.

Look at what the cited pages actually do. For the fee prompts, competitors' reviewed fee explanations are cited repeatedly and Horizon has no dedicated fee page at all. That is a coverage gap with an obvious response, and it is a response worth making because clients ask about fees constantly, not because publishing forces a citation.

The opposite pattern needs a different response. When your page is cited and the answer still describes you wrongly, the source is being read and misused, or the source itself is unclear. That is an editing problem on a page you own, which is the cheapest fix available in this module.

Choose the smallest verification step

Every diagnosis here is a hypothesis, and a hypothesis predicts what should change if it is right. Before commissioning work, ask what the cheapest thing is that would confirm or kill it.

For a suspected instability, the cheapest test is running the same prompt three more times today. For a suspected coverage gap, it is checking whether the engine cites anybody at all for that question. Both cost minutes and both routinely prevent a month of content work aimed at the wrong cause.

Publish because the audience needs it. The reviewed fee resource is worth building on its own merits. Framing it as a citation play sets an expectation nobody can guarantee, and it will be judged against that expectation rather than against whether it helped a single client understand what they would pay.

The gap that is not yours

Some absences are correct. A prompt asking who handles class actions in Delaware should not name an Austin personal injury firm, and an engine that leaves you out of it is working. Every prompt set contains a few of these and they quietly drag the mention rate down without indicating anything.

The signal that separates them is whether anybody comparable is named. When the answer lists four firms and none of them are you, that is a gap. When it lists nobody from your category at all, the prompt is measuring something else and belongs out of the cohort.

Third-party sources are a different remedy. Competitors named through directories, review sites, or press coverage you are absent from is not a content problem on your site, and writing another page will not address it. That is module 7's work, and misdiagnosing it as a coverage gap is how a team publishes six pages and moves nothing.

Diagnose one absence

Four steps, and the first two are reading rather than doing. Most absences are diagnosed without touching the site.

1. Read the full response and its cited sources.Not the classification. The sources name what the engine considered adequate, which is the most direct evidence available about what is missing.
2. Compare competitor mentions and their source domains.Whether they are cited through their own pages or through third parties. Those are two different problems with two different remedies, and module 7 owns the second.
3. Check your own coverage and consistency.Whether a page exists at all, and whether your pages agree with each other. Contradiction between your own pages is a common and entirely self-inflicted cause.
4. Pick the smallest step that would test the hypothesis.Usually a rerun or a check rather than a content project. Confidence should be established before spend, not after it.

Once a protocol is running, the tempting move is to reduce it to one number and put that number in a report. This lesson is about which numbers are real, what they can be compared against, and the specific way a single score goes wrong.

The short version: track several rates separately, keep them away from rankings and revenue, and never let an unmeasurable component quietly count as zero.

Rates move independently and mean different things. A rising mention rate with flat citations and a false claim is not a win.

Four rates, tracked apart

Citation rate is the one worth watching hardest, because it is the only one of the four that depends on your own pages being good enough to reference. A mention can come from anywhere, including a competitor comparison that names you as the weaker option. A citation means something you published was worth pointing at.

What a composite score may and may not do

A blended score is defensible when its components are published and each one is genuinely measured. Named, cited, named first, and share of voice against tracked competitors, per engine, with the weights written down.

An unmeasurable component must be dropped, never zeroed. If sentiment cannot be assessed for an engine this period, the score for that engine is computed without it and says so. Treating the missing component as zero manufactures a decline out of an absence of data, and that decline will be investigated as though it were real.

Publish the weights wherever the score appears. A composite whose formula is not visible cannot be argued with, which sounds convenient and means nobody can tell you when it has started measuring the wrong thing.

Keep it separate from rankings and revenue

AI visibility is its own layer. It is not organic ranking, and connecting it to revenue requires exactly the evidence module 1 demanded of every other connection.

The honest presentation puts it beside the other layers rather than inside them, with its own protocol, its own cohort, and its own dates. Horizon's tracker makes this concrete: referral traffic by AI engine sits on the same screen, and when analytics is connected but no traffic has synced, it says so instead of drawing a zero.

Attach the work and report the uncertainty

Record what shipped during each period alongside the movement, with dates. Not as proof, but so the plausible explanations are visible when somebody asks what caused a change.

Then state the confidence and the next verification date. A number reported without either is asking to be treated as more solid than it is, and in this part of the discipline that is the specific way credibility gets spent.

Report a tracking period

Six steps. The output should be legible to somebody who has never opened the tracker.

1. Confirm the cohort and protocol were stable.Same prompts, same conditions, no version break. If any changed, the period is two segments and should be reported as such.
2. Report the rates separately.Mention, citation, accuracy, competitor presence. A composite may accompany them and should never replace them.
3. Have a person review accuracy.On the material claims. This is the one that cannot be automated and the one whose failure costs the most.
4. Compute any score with its components visible.Published weights, and unmeasurable components dropped rather than zeroed. A score that hides a missing input reports a decline that did not happen.
5. List the work shipped in the period.With dates, as context rather than as attribution. Module 1's join labels apply here without modification.
6. State confidence and the next verification date.Including the alternative explanation you could not rule out. Then actually run the verification, which is what separates a program from a report.