Skip to main content
RankRift
Menu
Reporting & Impact

The AI Visibility Maturity Model

Ad-hoc, tracked, instrumented, attributed — four stages with entry criteria and the honest cost of each.

maturity-modelframeworkstrategy

Somebody senior asks whether you are doing enough about AI visibility, and you have no way to answer, because "enough" has no reference point. There is no published standard for how much a company of your size should be investing in measuring whether assistants name your brand.

This is an attempt at one. Four stages, each with entry criteria, a description of what it produces, and an honest account of what it costs. The point of the model is not to move you up it. For a lot of organisations the correct stage is the first one, and the model exists partly to make that a defensible answer rather than an embarrassed one.

Conflict note up front: RankRift publishes vendor-neutral frameworks about when tracking software is warranted. A maturity model published by a vendor is structurally suspect, so read this one adversarially. Our position on the conflict is on the disclosure page, and stage 1 below carries no product recommendation at all because it should not.

Stage 1: ad-hoc checks

What it is. Somebody in the team runs a handful of questions through a few assistants when the topic comes up, looks at whether the brand appears, and mentions it in a meeting. No fixed prompt set, no schedule, no stored data.

Entry criteria. Curiosity and an hour. That is genuinely all.

What it produces. Qualitative awareness. You learn roughly how your category gets described, which competitors surface, and whether your brand appears at all when somebody asks a buying question. That is real information and it is often enough to make the only decision that stage 1 is meant to inform, which is whether anything here warrants more attention.

Cost. A few hours per quarter of one person's time. No money.

What it cannot do. It cannot detect change. Because outputs vary between runs, an ad-hoc check that shows your brand today and not next month tells you nothing about whether anything changed. It cannot support a claim, a target, or a budget request, and it should never be put in a chart.

Staying here is a valid outcome. This is the part of the model that matters most. Many organisations should stay at stage 1 indefinitely, and the honest signal is straightforward: if you would not change any decision based on a measured movement in AI visibility, measuring it precisely is a cost with no return. A local services business with one location, a company selling through a single enterprise sales relationship, a team whose buyers do not research online, or an organisation whose category is barely represented in assistant answers at all, are all better served by an occasional look than by a programme. Choosing stage 1 deliberately, and writing down why, is a more mature position than drifting into stage 2 because a competitor mentioned they were tracking.

Stage 2: tracked

What it is. A frozen, versioned prompt set, run on a schedule against named models, with results stored and a provenance record maintained. The output is a time series with a stated spread rather than an impression.

Entry criteria.

  • A written prompt set that reflects how buyers actually ask, with a version identifier.
  • A defined primary metric with a formula.
  • Access to official model APIs, and a documented parameter set.
  • Someone accountable for running it on schedule.
  • Storage for raw responses.

What it produces. The ability to say whether a number moved, and whether the movement is distinguishable from run-to-run variation. That is the first genuinely defensible thing in this progression. It supports internal reporting and honest competitive comparison.

Cost. Honest ranges are hard to give without knowing your prompt count and model count, and anyone quoting a single figure is guessing. What is knowable is the shape: setup is a day or two of thinking and a day of building, running is API charges that scale with prompts times models times runs per cell, and maintenance is a few hours a month plus the cost of keeping a spreadsheet or script working when a provider changes something. Model API pricing is published per provider and changes; check current provider pricing pages rather than any figure quoted in an article. The recurring cost people underestimate is not compute, it is attention. A tracked programme that nobody maintains silently becomes an untracked one within a quarter.

Where to start. The provenance discipline is the whole of stage 2. Our guide to setting a defensible baseline is the entry procedure for this stage, and doing it in a spreadsheet first is a legitimate and often correct choice, covered in build, buy, or spreadsheet.

Stage 3: instrumented

What it is. Stage 2 plus the surrounding context: server-log evidence of AI crawler activity, referral traffic from assistant surfaces where it is identifiable, and an annotation log recording what your team changed and when.

Entry criteria.

  • A stable stage 2 programme running for long enough to know your own noise floor.

  • Access to server logs or a CDN log pipeline, and someone able to work with them.

  • Analytics configured to identify and segment assistant referrals, with an explicit acknowledgement of what that segmentation misses.

  • A discipline of writing down every material change on your own site, and observed changes on competitors', with dates.

What it produces. Context for the numbers. When visibility moves, an instrumented programme can look at whether crawler activity changed, whether referral traffic moved with it, and what your team shipped in the preceding weeks. It converts "the number went up" into "the number went up, and here are the three things that happened around then".

Cost. Materially higher and mostly in skills rather than in tools. Log analysis needs someone comfortable with log analysis. Referral segmentation needs analytics ownership. The annotation log needs a habit that survives busy quarters, which is harder than it sounds and is the component most likely to fail. Expect this stage to consume a meaningful fraction of one person's role rather than a few hours a month.

What it cannot do. It produces correlation and context, not proof. An annotation log tells you what happened at the same time. It does not tell you what caused what.

Stage 4: attributed

What it is. Visibility measurement connected to pipeline: leads and revenue tagged in a way that lets you say something about the commercial value of appearing in AI answers.

Entry criteria.

  • A working stage 3 programme.
  • A CRM with reliable source capture and a sales team that fills it in.
  • Analytics and revenue-operations capability, plus someone who understands the statistical limits of attribution and is empowered to say so.
  • A leadership team that will accept a range and a caveat rather than a single number.

What it produces. Directional commercial evidence. At best, a defensible statement that a segment of pipeline appears to be associated with assistant-driven discovery, with the uncertainty attached.

Cost. The highest, and the least predictable. The measurement work is the small part; the expensive part is the organisational work of getting source data captured consistently at every touchpoint.

The hard limits nobody has solved. Be blunt with yourself about this before committing budget to stage 4.

  • Assistant-influenced discovery frequently leaves no trace. A buyer who reads a recommendation in an assistant and then searches for your brand name arrives as branded search. That path is invisible to source capture, and the volume of it is unknown.
  • Zero-click behaviour means influence can occur with no session at all.
  • Self-reported source fields in a CRM are unreliable, and their unreliability is not random.
  • There is no experimental control available. You cannot randomise which buyers see your brand in an assistant answer, so the strongest available evidence is observational and confounded.

None of that makes stage 4 worthless. It makes stage 4 a stage whose output is a carefully hedged range, and any vendor or agency offering clean AI-visibility revenue attribution is selling something the underlying data cannot support. If reaching stage 4 requires you to promise precision you cannot deliver, staying at stage 3 is the more credible position.

Entry criteria and cost at a glance

StageEntry criteriaProducesCost shapeMain risk
1. Ad-hocCuriosity, an hourQualitative awarenessHours per quarter, no spendMistaking variation for change
2. TrackedFrozen prompt set, defined metric, API access, an ownerA defensible time series with spreadSetup days plus recurring API spend and monthly maintenanceQuiet decay when nobody maintains it
3. InstrumentedStable stage 2, log access, analytics ownership, annotation habitContext and correlationA significant fraction of one roleCorrelation presented as cause
4. AttributedStage 3, reliable CRM source capture, analytics capability, tolerant leadershipHedged commercial estimatesHighest and least predictable, mostly organisationalOverclaiming on data that cannot support it

How to decide which stage you should be at

Work through these in order and stop at the first honest no.

  1. Would a measured change in AI visibility change a decision you make? If not, stay at stage 1 and revisit in six months.
  2. Do you have someone who will own a schedule? Without an owner, stage 2 will not survive a busy quarter, and a decayed programme is worse than an admitted absence because it produces numbers nobody is maintaining.
  3. Do you need to defend a number to someone with budget authority? That requirement, and not company size, is what justifies stage 2.
  4. Do you have log and analytics capability in-house or reliably available? Without it, stage 3 is aspiration.
  5. Will your leadership accept a range with caveats as an answer? If the only acceptable output is a clean revenue figure, stage 4 will produce a dishonest one. Decline it.

Two observations from how these decisions usually go wrong. Companies skip stage 2 and attempt stage 3 or 4 because attribution is what leadership asked for, then produce context and correlation on top of a measurement base that was never stable. And companies at stage 2 keep adding prompts and models until maintenance collapses, when the more useful move is holding the set frozen for longer. Depth of series beats breadth of coverage for almost every question you will be asked.

Nothing in this model implies you should end up buying anything. A spreadsheet at stage 2 that is actually maintained is worth more than a platform at stage 3 that nobody reads. If your prompt count is small, your models few, and your team comfortable with a scheduled script, doing it yourself is the right call and will stay the right call for a long time. Tooling starts to earn its cost when the maintenance burden, the number of brands, or the reporting obligation exceeds what a person can hold in a spreadsheet, and not before.

In Plain English

Four stages: ad-hoc checks, tracked measurement, instrumented context, attributed pipeline. Each has real entry criteria and a real cost, and the costs rise faster than the certainty does. Stage 1 is a legitimate destination, not a starting insult, and the test is simply whether a measured change would alter any decision you make. Stage 4 has limits that no method currently solves, so its honest output is a hedged range rather than a clean number.

This helps anyone being asked whether their level of investment is right: an in-house lead sizing a programme, an agency setting client expectations, or a founder deciding whether this deserves attention yet. Use it before you commit budget, and use it again when someone proposes jumping two stages at once.

Next: identify your current stage, decide with reasons whether moving is warranted, and write that decision down, including the case for staying put. The reporting hub covers what each stage lets you legitimately say out loud.

If you have decided stage 2 is right for you, start tracking with a frozen prompt set and a provenance record, whether in a spreadsheet or in a tracker. If you have decided stage 1 is right for you, that decision needs no product at all.

Primary-source verification

Checked July 28, 2026. The operational and factual guidance above was reviewed against the primary references below. Provider features, prices, policies, interfaces, jurisdictions, and account-specific results can change; recheck the controlling first-party document and your own account or measured data before acting.