Skip to main content
RankRift
Menu
Getting Recommended

A First-90-Days AI Visibility Plan

Baseline, levers, re-measure — a sequenced plan that also names what to skip in the first quarter.

planningroadmap90-days

The approval came through and now you need a sequence you can defend in a review. The temptation is to fill the quarter, because a full plan looks like seriousness. A full plan is how measurement gets dropped in week nine.

This is a thirteen-week sequence with four workstreams and a deliberate skip list. Every action below names how it will be measured, because an action with no measurement attached is just an opinion with a deadline. Read the skip list before the plan; it is the part that makes the rest fit.

Set expectations first. A quarter is enough time to establish a defensible baseline, ship two or three genuine changes, and re-measure honestly. It may not be enough time to see those changes reflected in assistant answers. Model update cycles, retrieval refresh, and third-party publication timelines are all outside your control and can run longer than thirteen weeks. Plan the quarter as work completed and measurement established, not as movement achieved. The section on what would change the plan covers what to do when week thirteen arrives flat.

Weeks 1–2: baseline and provenance

Nothing else in the quarter is interpretable without this. Do not compress it.

  • Define the prompt set. Twenty to forty prompts a real buyer in your category would type, covering the category question, the comparison question, the constraint question, and the alternative-to question. Write the selection rule down. Measured by: the prompt set is versioned and stored, so week eleven runs the identical set.
  • Pick two or three models and pin the versions. Use official model APIs. Record model identifiers and dates on every run. Measured by: each result row carries model, version, and run date.
  • Decide your sample size before running. How many runs per prompt per model, chosen deliberately rather than by budget accident. Measured by: stated sample size in the baseline record.
  • Run the baseline and record variance, not just averages. You need the spread across runs to know what counts as a real change later. Measured by: a baseline table with mention rate, citation share, and run-to-run variation per metric.
  • Write the provenance block. Prompt set version, models and versions, run dates, sample size, variance. Measured by: any colleague can reproduce your numbers from the block alone.
  • Set the action threshold. How much movement would justify changing plan. Decide now, while you have no result to be tempted by.

The mechanics of doing this so it holds up six months from now are covered in setting a defensible baseline. If you cut anything from weeks one and two, cut prompt-set size, never variance.

Weeks 3–6: the two highest-evidence levers

Two levers. Not five. Two chosen because they are the ones with the clearest mechanism between the work and what an assistant can retrieve about you.

Lever one: entity clarity on your own properties. Make it unambiguous, in plain text, who you are, what category you are in, what you sell, at what tier, for whom, and who you are not. This includes consistency of your name and descriptors across your own site and the profiles you control, and removing the internal contradictions that accumulate when six people write about the same product over three years. Entity clarity covers the specifics.

  • Measured by: a documented before-and-after of the descriptor set, plus a re-run of the subset of baseline prompts that ask "what is X" and "who is X for." Those are the prompts most likely to reflect this work first.

Lever two: retrievable answers to the questions your prompt set asks. Where your baseline shows assistants answering a buyer question without you, check whether a clear, self-contained answer to that exact question exists anywhere you control. Often it does not, or it exists buried in a page about something else. Write the direct answers. Structure them so a single section stands alone.

  • Measured by: a mapped list of baseline prompts to the page that now answers each, and the count of prompts with no corresponding page (which should fall).

Budget the full four weeks. Both levers are writing and editing work, and both are done badly at speed. Resist adding a third lever with the four weeks you think you have spare. You do not have them.

Weeks 7–10: source and coverage work

This is the slowest-returning workstream and the one most likely to slip past the quarter. Start it now precisely because it will not finish.

  • Identify the sources actually cited in your category. From your baseline runs, extract which domains and publication types appeared in citations across models. This is a data exercise, not a guess. Measured by: a ranked source list with counts and run dates attached.
  • Assess your presence in each. For each recurring source, record whether you appear, appear inaccurately, or are absent. Measured by: a three-state audit table you can re-run next quarter.
  • Pursue the two or three most achievable gaps. Accurate listings, corrections to outdated descriptions, legitimate inclusion in relevant roundups and directories that do not sell placement. Earning cited sources covers what legitimately works and what to refuse.
  • File corrections where a source states something false about you. This is often the highest-value item on the list and the one nobody does, because it is unglamorous.
  • Measured by, across the workstream: count of gaps closed, with dates, and a note of which are expected to affect retrieval and which are simply accuracy hygiene.

Expect zero visibility movement from this workstream inside the quarter. Third-party editorial timelines and model retrieval refresh both run on schedules you do not set. The measurement here is work completed, not visibility gained. Say that in the plan document so nobody is disappointed at week thirteen.

Weeks 11–13: re-measure and report

Protect these three weeks. This is the part the quarter exists for.

  • Re-run the identical prompt set against the same models, recording current versions. If a provider shipped an update in the window, note it; it is a candidate explanation for anything that moved.
  • Compare against the baseline using your variance figures. A change smaller than your run-to-run variation is not a change. Report it as such.
  • Compare against your action threshold, not against zero.
  • Write the report. Baseline, work completed, re-measurement, honest verdict, what you would do next and why. Include the provenance block.
  • State what you cannot claim. You can show that you did the work and what the metric did. You cannot attribute the metric to the work from a single before-and-after with no control. Say so plainly; it is what makes the parts you do claim believable.
  • Measured by: the report exists, is reproducible, and reaches whoever controls the next quarter's budget.

If week eleven arrives and the levers are unfinished, re-measure anyway. A clean measurement of incomplete work is more useful than a delayed measurement of complete work.

What to deliberately skip this quarter

The skip list is not a list of things that do not matter. It is a list of things that do not fit alongside a defensible baseline in the same thirteen weeks.

Skip this quarterWhy
Buying a tracking platformYou do not yet know your prompt set, cadence, or which metrics you will actually report. Buying first means buying against a guess and then shaping your method around whatever you bought.
Tracking more than two or three modelsEvery added model multiplies run cost and analysis time and adds nothing to your first read. Add breadth once your method is stable.
A large competitor benchmarkCompetitor comparison is seductive and expensive, and it does not tell you what to change. Establish your own baseline first.
Site-wide technical or schema overhaulsHigh effort, slow feedback, and hard to attribute against everything else moving. Fix crawler access if it is genuinely broken; otherwise defer.
Chasing every prompt where you are absentThe absence list is long by design. Working the whole list dilutes both levers into nothing.
Weekly reportingWeekly readings on a metric with this much run-to-run variation produce noise that people will react to. Monthly at most, quarterly for leadership.
Building attribution from assistant answers to revenueThe pathway is genuinely hard to observe end-to-end and the plumbing work is a quarter of its own. Start with referral detection only, and be honest that it is partial.
Any tactic that manufactures signalsAstroturfed community posts, incentivized reviews, and content networks designed to be scraped are liabilities rather than levers, and detection costs land on the brand rather than on whoever sold the tactic. They are not on this plan at any speed.
Publishing a "we measured the category" studyOriginal research is a strong asset and a poor first project. Your method needs one honest internal cycle before it goes public.

Write your version of this table into the plan document and get it acknowledged in the same meeting that approves the plan. A skip list nobody agreed to is just work you failed to do.

What would make you change the plan

Decide the triggers now, so week thirteen is a decision rather than an argument.

  • The baseline shows near-total absence across every prompt and model. Then entity clarity is not the constraint; basic retrievability may be. Check crawler access and rendering before spending four weeks on writing.
  • The baseline shows high variance and low mention rates. Your sample size is likely too small to detect anything you will do. Fix sampling before choosing levers, or accept you are running a work-completed quarter with no readable measurement.
  • A provider ships a significant change mid-quarter. Re-baseline against the new version rather than comparing across it. Note the discontinuity in the report; do not paper over it.
  • A source states something materially false about your brand. That jumps the queue ahead of everything in weeks three to ten.
  • Week thirteen is flat. This is the expected case as often as not. Flat after one quarter is not evidence the levers failed; the observation window may simply be shorter than the effect. Run one more cycle before changing approach, and say in advance that you will, so a flat quarter does not read as failure to leadership.
  • Week thirteen moves a lot in either direction. Check for a model update before crediting or blaming yourself. Large sudden movements are more often the system than you.

The broader question of which levers are worth the second quarter is covered in what actually changes recommendations.

In Plain English

Spend two weeks building a baseline you can reproduce, four weeks on exactly two levers, four weeks starting slow source work that will not finish, and three weeks re-measuring and reporting. Skip the tool purchase, the competitor benchmark, the model breadth, and weekly reporting until your method is stable. Treat the quarter as a measurement program you established rather than a result you delivered, and expect that thirteen weeks may be shorter than the time it takes for anything to show up in assistant answers. This is for teams with approval to start and a review to survive. Next step: write your prompt set and your skip list this week, in that order.

When you are ready to run this on a schedule rather than by hand, start tracking with the tracker.

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.