Most 90-day GEO plans end in a slide that says visibility is up, built on three prompts run twice. That is reporting noise as progress: SparkToro's research found AI engines are highly inconsistent when recommending brands, and a 2026 variance-components study found run-to-run variation large enough to swamp real differences in small samples. A quarter is enough to finish three things: a baseline you can rerun, a site AI crawlers can read, and extractable proof on the pages your buyers ask about. Reliable movement in what a model recommends is not one of them.
The work is still worth doing on that clock. G2's research found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers now start research with an AI chatbot more often than with Google, and a third of buyers said an AI chatbot led them to choose a vendor they hadn't considered. 6sense found buying groups consider roughly five vendors, leaving about one spot open for a vendor they have not had prior experience with. What follows is a dated sequence built so that day 90 produces an answer you can defend.
What can 90 days of GEO actually change?
Three things, each on a different clock. Retrieval fixes land fastest, because crawlers re-fetch on a schedule you do not control; the volume of that crawling is climbing too, with Cloudflare tracking GPTBot's share of AI and search crawler traffic going from 2.2% to 7.7% over a year. Evidence added to your own pages moves next, as those pages are re-crawled. Presence inside the third-party sources engines cite moves slowest of the three.
What does not move on a 90-day clock is the association a model already carries for your category. A 2026 analysis of brand dynamics in LLM recommendation systems found well-known brands were recommended every time when the products had identical specifications, though the edge disappeared once a competitor had less than a +0.1-star rating advantage. In one quarter you are competing for the retrieval slot and the quality of the evidence inside it.
That sets the honest ceiling on the plan. Google states that its AI features carry no additional requirements or special optimizations, need no new machine-readable files or markup, and that pages must be indexed and snippet-eligible to appear at all. The first quarter is mostly about clearing gates you can verify yourself, which is why the sequence below is deliberately boring.
What goes in days 1 to 30?
One deliverable: a baseline specific enough that day 90 can contradict it. Write down 15 to 25 questions a real buyer asks before choosing in your category, covering category questions, problem questions, comparisons, and your own brand name. Record five fields per run: mentioned, cited, recommended, shortlisted, and which sources carried the answer. Shortlisted means you survive the narrowing when the answer compares options, which is a different event from being named in it.
Ask each question in two or three phrasings, on at least two engines, every week. The phrasings and the weeks are not padding. Search Engine Land's write-up of the SparkToro and Gumshoe study reports the same prompt rarely returns the same vendor set twice, and Sielinski's work on quantifying uncertainty in AI visibility puts confidence intervals around exactly this kind of sample.
The same month covers retrieval, because measuring pages an engine cannot read wastes the measurement. Vercel's crawler research with MERJ documented that GPTBot and its peers read raw HTML rather than executing JavaScript, and Cloudflare's agent-readiness scan of the 200,000 most visited domains found the standards built for agents barely adopted: only 4% of sites declared AI usage preferences in robots.txt, and 3.9% passed Markdown content negotiation. Fetch your ten most important pages with JavaScript off and read what survives.
Close the month by confirming you block nobody by accident. OpenAI documents four relevant user agents, each with a different job, and one careless line in robots.txt can cut you out of a channel you were counting on. That check takes an hour and it gates the other 89 days.
What goes in days 31 to 60?
Close the specific gaps your baseline named, in the order the baseline ranked them, before anyone opens a content calendar. Mechanical problems go first because they are cheap and verifiable. Everything else queues behind them, and source problems will not finish inside this window.
Evidence work means replacing assertion with something a model can extract and attribute. The GEO study (Aggarwal et al., KDD 2024) found that adding statistics, quotations, and citations lifted citation visibility by roughly 30 to 40% in its benchmark, while its own error bars showed keyword stuffing doing nothing. Some persuasion language actively costs you: scarcity and exclusivity framing measurably reduces how often a model recommends a product.
Profound's citation research found more than 4 in 10 AI citations point to sources brands do not control: editorial coverage and community threads. If your baseline shows answers assembled entirely from third-party pages that omit you, nothing published on your own domain fixes that by day 60. Start the earned work anyway and book it as a month-four outcome.
Name the assets now, ranked by what the day-30 record shows and not by a calendar. The comparison page usually goes first, because no third-party source will write it for you and a narrowing answer has nowhere else to pull it from. The proof page goes next: the benchmark or priced example that answers the question a rival's quoted evidence currently owns. Third, only if the first two land early, is the buyer-question page for the question you appear in least.
Five hours a week is about twenty hours inside this window. That is two assets built to the standard above, which is why the third one is conditional. AirOps' March 2026 analysis of 548,534 pages ChatGPT retrieved found only 15% were cited in the final answer, so each page has to earn selection, not just retrieval. Publishing cadence is a separate variable, and twenty hours does not cover both.
What goes in days 61 to 90?
The rerun: identical questions, identical engines, identical run counts, compared against the day-30 record. Changing the question set between runs is the most common way teams destroy the experiment. If a question turns out wrong, keep running it and add the better one as a second series.
Read appearance rates. The platform medians in that same uncertainty work show how wide a plausible range sits around any one measurement, so read a move on one question in one week as noise, and a move that holds for several weeks across a group of questions as signal.
Score the rung rather than the mood. Mentioned, cited, recommended, and shortlisted are four outcomes with four different causes, which is why the IAB's August 2026 framework splits brand visibility into Presence, Prominence, Portrayal, and Persuasion. A composite number would bury the one distinction you can act on.
Look at traffic last, and expect it to be thin. GA4 defines session key event rate as sessions with a key event divided by total sessions. On a cohort this small that rate swings on single sessions, so record absolute counts now and wait for volume before reading the rate at all.
In its December 2025 re-run, Ahrefs measured a 58% lower average clickthrough rate for the top-ranking page on informational keywords where an AI Overview appears. Thin click volume from AI surfaces is the ambient condition, so do not read it as a verdict on your quarter.
Which parts of the standard 90-day plan should you skip?
Skip any tactic whose only evidence is an article recommending it. The clearest case is prose rewriting. C-SEO Bench (NeurIPS 2025) tested ten conversational-SEO methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. Restyling existing copy to sound more quotable is the least reliable line item in most 90-day plans.
Skip the single visibility score as well. There is no shared definition of its denominator, the engines disagree with themselves between runs, and a composite figure moves for reasons you cannot trace back to a decision. Track the four rungs separately and accept a messier dashboard.
Do not skip organic search on the theory that it is finished. Ahrefs found 38% of AI Overview citations rank in the organic top 10 across 4 million AI Overview URLs in March 2026. Rankings alone do not carry you, and 38% means they are still the largest share of AI Overview citations a startup with no earned coverage can reach.
The same dependency shows on another surface: Seer's 500-citation check of SearchGPT answers found 87% matched Bing's top results. Google's helpful content guidance still describes what earns that position.
How does Trovance run the baseline, gap, and rerun cycle?
Trovance runs this 90-day sequence as a standing cycle instead of a project with an end date. You define the tracked questions your buyers actually ask, and weekly sweeps run them across engines. Each answer snapshot preserves the full context: who was mentioned, who was cited, who was recommended, who survived the shortlist, and which sources carried the answer. That record is your baseline, and it keeps accruing past day 30.
The gap work then comes from comparing snapshots. Share of answer shows which buyer questions you appear in and which you are absent from, and the preserved citations show whether a competitor won on their own page or on someone else's. The Brain keeps the claims you are entitled to make with the proof behind each one, and the ranked queue, backed by answer evidence, names a specific asset: the benchmark the answers keep quoting from a rival, or the comparison page that does not exist yet.
Day 90 is a rerun inside the same system. The analysis cycle reruns the same tracked questions after your work ships and compares the new snapshots against the old, so you can observe whether the rung moved. Drafts are written in your voice from your voice guidance and research sources, a person edits and approves everything before it publishes, and the next cycle verifies what actually shipped.
Trovance will not promise that a citation or a recommendation appears by day 90. The engines are probabilistic, your competitor is publishing too, and no single visibility score exists that would make such a promise measurable even if it were honest. What the cycle gives you is a dated record of what the answers said before and after your work, which is the only sound basis for planning days 91 to 180.
What should you do this week?
Write the question list today and run it tomorrow. Fifteen buyer questions, two or three phrasings each, two engines, weekly, five fields recorded per run. That is the entire first week, and it is what makes the next 83 days falsifiable.
Then fetch your five most important pages with JavaScript disabled, read what an engine would read, and fix what is missing before writing a word of new content. Put the day-90 rerun in the calendar now with the question list attached, so nobody has to reconstruct the method from memory in three months.
Be honest in the readout you promise your board: retrieval fixes can surface once engines re-fetch the page, evidence changes follow re-crawling and re-retrieval, earned sources take longest, and no plan guarantees a citation on a date. If you would rather run the baseline and the rerun as a standing cycle than as an afternoon you keep rescheduling, start a free Trovance trial and let day 90 compare itself to day 1.
Set the baseline
The 10-minute audit of what ChatGPT tells buyers about you - the fastest first reading of where you stand.
Build the prompt library you will rerun - the question set day 90 compares against.
How to measure AI search visibility without one score - what to record instead of a composite number.
Technical SEO for AI search - the retrieval work that belongs in month one.
Close gaps and rerun
Run an AI citation gap analysis - turning a baseline into a ranked list of gaps.
A visibility gap is not a content brief - why not every gap earns a page.
How long until AI cites you - what honest timing looks like across the rungs.
Measuring GEO return without fooling yourself - how to read the second run.
