A 90-Day GEO Program: How We Structure Engagements
How our agency adapts a 90-day GEO framework to establish a baseline, ship focused work, measure movement, and set an ongoing operating rhythm.
Ninety days is a useful planning window for a GEO engagement, but it is not a scientific law or a promise that visibility will improve on a schedule. Pages need to be accessible, discovered, considered as sources, and reflected in generated answers. The timing varies by site and model. A fixed program helps us order the work and keep clients from judging a new measurement setup after a few days.
Our structure adapts Promptwatch's 90-day GEO guide. That guide assigns baseline and setup to month one, diagnosis and publishing to month two, then measurement and operating rhythm to month three. We use the sequence because it separates observation from intervention. We still change the work plan when the client's technical state, publishing capacity, or market calls for it.
Before day one
The engagement needs a narrow commercial scope. We agree which products or services matter, the markets and languages in scope, who the actual competitors are, and which teams can approve technical and editorial changes. We also identify the business actions that matter after an AI referral, such as a qualified form submission or completed purchase. Those outcomes stay in the client's analytics or CRM.
Access comes next. We request the visibility platform, analytics, tag manager, content system, Search Console or sitemap, and the relevant CDN or log source. If legal or security review will delay a connection, we record that early. The program clock should not conceal a measurement dependency.
Month one: build a baseline
The first month is for setting up a trustworthy observation set. We confirm the brand name, domain, aliases, products, and comparison set. Then we build prompt groups from customer language and buying situations, not a pile of keyword variants. Each prompt has a reason to exist: it tests category discovery, comparison, use case, reputation, or direct brand understanding.
Model and location choices are explicit. Adding every available model can consume response capacity without improving the decision. We select models that fit the audience and keep one market per monitor when location differences matter. Topics and tags make later analysis possible without rebuilding the prompt set.
We connect page inventory through a sitemap or Search Console and establish crawler-log and visitor measurement where the stack allows it. The crawler check comes early because a blocked page cannot compete for a fresh citation. Promptwatch's crawlability guide distinguishes training crawlers, search index crawlers, and live citation fetchers. We audit those purposes separately and check each provider's current official documentation before recommending an allow or block rule.
During the remaining weeks, we correct setup errors rather than launching a broad content sprint. A missing alias, wrong country, irrelevant competitor, or vague prompt can contaminate the baseline. At the end of the month, we record visibility, share of voice, answer position, sentiment, citation measures, crawler status, and identified AI referral visits. The Promptwatch metrics guide supplies definitions for its own metrics. We label them as platform measures rather than universal scores.
Month two: diagnose and ship
With a stable observation period, we look for specific losses. We split results by model, topic, prompt, competitor, and cited source. A blended average is rarely enough. One model may know the brand but cite third parties, while another may omit it from category prompts. Those are different jobs.
Page-level diagnosis follows the evidence chain. No crawler requests suggests a discovery or access problem. Successful crawls without citations point toward source selection, content coverage, or authority. Citations without a brand mention can mean the page supplies a fact while the answer favors another company. Mentions without self-citations suggest that outside sources shape the model's description.
We turn that diagnosis into a limited delivery queue. Technical blockers are assigned to the infrastructure owner. Existing pages with a clear coverage gap get revised before we commission a new library of articles. New pages are mapped to prompts and buyer needs, with an owner and publication date. Publishing steadily gives earlier work more time to be fetched and observed before the end of the program.
Not every gap belongs on the client's website. If tracked answers repeatedly cite community threads or videos, the work may involve genuine participation, creator outreach, or useful video production. We use the Reddit and YouTube strategy guide as product guidance for investigating cited communities and channels. We do not treat it as permission for undisclosed promotion. Community rules and honest affiliation come first.
Month three: measure and establish a rhythm
The third month compares work with the month-one baseline. We track published and revised pages individually where possible, then look for the sequence of successful crawl, citation, and identified click. The absence of one step changes the next action. More rewriting will not fix a 403, and a firewall change will not make a weak page worth citing.
We also review movement at the prompt-group and model level. A change concentrated on prompts served by a revised page is more informative than a broad swing across the whole project. Even then, we describe association rather than claiming certain causation. Generated answers can shift because sources, retrieval systems, and models change outside the engagement.
The day-90 report includes the original scope, baseline, work shipped, technical issues resolved or open, observed movement, identified AI referral traffic, and the next quarter's queue. Revenue attribution comes from the client's own systems. Referrer-based measurement is a floor because some visits arrive without a referrer and some people return through direct or branded routes.
We use Promptwatch for client programs because one project can connect prompt observations with citations, crawler activity, page tracking, and identified AI referral visits. The platform does not remove the need for editorial judgment or client analytics. It gives our team a common evidence base.
After day 90, the useful result is a repeatable cadence: review the data, diagnose a bounded set of gaps, ship approved work, and inspect what happened. We retire prompts that never inform a decision and add new ones only when buyer language or the client's offer changes. The program finishes with a working queue and named owners, not a claim that GEO is complete.