The GEO flywheel: plan → produce → publish → measure

Here's a pattern we see constantly. A team discovers their brand is invisible in AI answers. Leadership funds a "GEO project": an agency audit, a batch of page rewrites, a schema rollout, a triumphant slide. Six months later someone checks again — and the numbers have drifted back, or a competitor has overtaken them, and nobody can say why, because nobody was looking.

The failure isn't the work. It's the shape of the work. Search behavior is migrating to assistants on a scale Gartner pegged at a 25% decline in traditional search volume by 2026² — and the systems assembling those answers retrain, re-index and re-rank continuously. A one-time optimization of a moving system is a snapshot, not a strategy. What survives contact with a moving system is a loop.

The flywheel

Four stages, each with a single job and a concrete output that feeds the next:

1. Plan — turn measurement gaps into a work queue

The plan stage owns two artifacts. First, the prompt suite: 20–50 prompts phrased the way your buyers actually ask, spanning discovery, comparison and problem-first intents. This is your instrument; it changes rarely and deliberately, because every change breaks comparability. Second, the priority list, derived directly from measurement gaps: prompts where you're absent, engines where you lag, claims answered wrongly, competitors over-indexing. Planning that doesn't start from measured gaps is guessing with a roadmap template.

One practical note on ranking: weight gaps by commercial intent, not by how winnable they look. Being absent from "best [category] for [your best-fit segment]" is worth more than being under-cited on a broad informational prompt, even if the second is easier to move. The prompt suite should encode this — tag each prompt with the revenue motion it represents, and let those tags drive the queue.

Output: a ranked list of gaps, each mapped to a content or corroboration action.

2. Produce — build pages engines can use as evidence

Production takes the top of that queue and builds what's missing: evidence-first product pages, honest comparisons, FAQ sections mirroring real buyer questions. The research is unambiguous about what wins here — statistics, quotations and cited sources lifted source visibility by up to ~40% in controlled tests, while keyword stuffing did nothing.¹ The discipline is refusing work that doesn't trace to a gap: producing content because the calendar says so is how flywheels become hamster wheels.

Output: shipped pages where every commercial claim carries a number, a source, or both.

3. Publish — make the work retrievable

A perfect page that engines can't ingest is a draft. Publishing means the technical layer: server-rendered HTML, Product/FAQ schema in sync with visible content, robots.txt that admits the AI crawlers you want, an llms.txt manifest pointing engines at your canonical pages,³ and no specs gated behind PDFs or forms. This stage is cheap and mostly one-time per page — which is exactly why it's forgotten, and why a quarterly re-check belongs in the loop.

Output: pages verified crawlable and machine-legible, confirmed via server logs showing bot visits.

4. Measure — rates and intervals, not anecdotes

The measure stage runs the suite: each prompt, multiple times, on each engine, on a schedule. It produces mention rate, recommendation rate and share of voice — with confidence intervals, because single runs of stochastic systems are noise. It also tracks accuracy: what the answers say, not just whether you appear. The stage's whole purpose is to feed stage one: this quarter's measured gaps are next quarter's plan. That handoff is the loop closing — and it's the handoff most teams never build.

Output: a delta report against last period, decomposed into the gaps that become the next plan.

Why one-off projects plateau and loops compound

Three properties of the generative-answer environment make loops structurally better than projects:

1. The ground moves. Models retrain, indexes refresh, engines redesign retrieval. A fix validated in March may be irrelevant by September — but only a team that's still measuring notices in September.
2. Competitors adapt. Share of voice is zero-sum. When your comparison page starts winning, someone else's stops — and the losing side eventually responds. A project has no mechanism for the counter-move; a loop absorbs it as next cycle's gap.
3. Wins feed wins. A page that gets cited gets read, linked and paraphrased by other publishers — whose pages engines also retrieve, corroborating your claims into consensus. Each turn of the wheel starts from a higher baseline. Compounding is slow at first and embarrassing to bet against later.

There's also an organizational reason: loops survive personnel changes and budget cycles because they're a cadence, not a hero project. The audit-shaped approach depends on someone re-noticing the problem every year; the loop notices automatically.

It's worth naming the ways the wheel stalls, because they're predictable. The most common is the open loop: measurement happens, a report gets circulated, and nothing in the next production batch traces to it — stages one and four exist, but the handoff doesn't. Second is instrument churn: someone "improves" the prompt suite every month, which feels rigorous but destroys every trend line and makes quarter-over-quarter claims meaningless. Third is production theater: shipping volume against no measured gap because content velocity is the metric someone reports upward. Each of these looks like activity from the outside; the tell in all three cases is that nobody can connect this month's work to last month's numbers.

The cadence that makes it sustainable

The stages run at different speeds, and forcing them onto one calendar breaks the loop. A cadence that holds up in practice:

Note the deliberate asymmetry: measurement runs faster than production. That's what makes attribution possible — when you ship quarterly and measure monthly, you get three readings per production batch, enough to distinguish real movement from noise. Teams that ship weekly and measure quarterly have the ratio backwards and end up with beautiful content and no idea what worked. Running this loop end to end is, not coincidentally, exactly what MentionBeat is built around — the platform automates the measure and plan stages so your team's time goes into produce and publish.

Your first cycle, concretely

• Weeks 1–2: build the prompt suite from sales transcripts; run the baseline across engines with repeated sampling.
• Week 3: write the first plan — your top ten gaps, each with an owner and an action.
• Weeks 4–10: produce and publish the first batch: 3–5 rewritten money pages, schema, llms.txt, one third-party placement.
• Weeks 6, 10, 14: monthly measurement runs; compare against baseline with confidence intervals.
• Week 14: close the loop — turn the delta report into cycle two's plan.

The first turn of a flywheel is always the heaviest — you're building the instrument and the habit at once. The second turn is easier, and by the fourth, the question "is GEO working?" answers itself from a dashboard instead of a debate.

A final framing for whoever has to sell this internally: the loop is also the cheapest honest answer to "what's our AI strategy for marketing?" It doesn't require betting on which engine wins, which model ships next, or which tactic survives the year — the measurement stage tells you, continuously, against your own buyers' prompts. Strategies that depend on predictions age badly in this space. Strategies that depend on feedback don't age at all; they just keep turning.

Frequently asked questions