Page one of a monthly AI visibility report should carry exactly four numbers: presence rate with a confidence band, share of answers against a named competitor set, AI-referred sessions and conversions, and prompts won and lost this period. Everything else — the per-engine split, the citation sources, the methodology — belongs on pages two, three and the appendix. The reason page one has to be that disciplined is the same reason the report is hard to write at all: in the SparkToro and Gumshoe study of 2,961 queries run by 600 volunteers in late 2025, two runs of the same prompt on the same platform returned the same brand list less than 1 time in 100, and the same list in the same order less than 1 time in 1,000.
Which means your client's Tuesday-morning spot-check will disagree with your report. Not might — will. They open ChatGPT, type the prompt they know you track, see three competitors and no mention of themselves, and email you a screenshot at 9:14am. The monthly report either survives that email or doesn't, and the outcome is decided before it arrives, by how you framed the numbers the month before.
The report is where the GEO retainer is won or lost
Start with the commercial reality, because it shapes the format. AgencyAnalytics surveyed 494 agency professionals between February and April 2026 and found 69% report to clients monthly, up from 65% the year before, with only 11% reporting weekly. Ninety-seven percent said accurate reporting is important or extremely important to retention; 76% called it extremely important.
That reads like a mandate to report more. It isn't, quite. The same survey puts the top reason clients leave in 2026 at budget cuts and economic pressure (42%), with "lack of perceived value over time" at 32% and attribution or ROI-clarity issues at only 10%. So a better report does nothing about the 42% — no PDF survives a budget freeze. It defends the 32%: the slow erosion where a client stops being able to explain internally what you do for them.
Demand, meanwhile, is not in question. Sixty-six percent of agencies reported rising client demand for AEO and SEO for AI-driven search, the number one new service category, ahead of performance-based paid ads at 55%. Most of those clients have no framework for judging whether what comes back is any good, so whatever you put on page one in month one becomes the standard you're measured against for the life of the retainer.
Two format constraints follow. Eighty-four percent of clients rank clear visual data representation above everything else, and 46% of agencies build a report in under 30 minutes, 73% within an hour. A GEO report that eats two hours per brand across a dozen brands fails on economics long before it fails on accuracy.
Page one: the four numbers
| # | Number | What it answers | Where it comes from |
|---|---|---|---|
| 1 | Presence rate ± confidence band | "How often do the engines mention us?" | Frozen prompt set, rolling 4 weeks, all engines |
| 2 | Share of answers vs named competitors | "Are we winning or just present?" | Same prompt set, same denominator |
| 3 | AI-referred sessions and conversions | "Did any of this reach the business?" | GA4, Observed tier only |
| 4 | Prompts won and lost this period | "What changed, specifically?" | Movement above the stated threshold |
Presence rate, always with a band. The percentage of tracked answers mentioning the brand, across a frozen prompt set, over a rolling four-week window — never a single-run score. Kevin Indig's framing on Search Engine Land is the one to internalise: treat prompt tracking as polling rather than rank tracking, run each prompt about five times per platform per week, and report a rate ± a confidence interval. His example reads "HubSpot is mentioned in 78% ± 6pp of problem prompts on ChatGPT vs. 34% ± 9pp on Perplexity," and his verdict on the alternative belongs in your appendix: "Reporting a point estimate from one run is astrology."
The band is what makes a drop survivable. A client who has seen "62% ± 5pp" for three months reads 58% next month as weather. A client who has only ever seen "62" reads 58 as four points of their fee going missing.
Share of answers against a named competitor set. Presence rate alone can't tell a client whether they're gaining, because category visibility moves with the category — when Google changes a model, everyone's numbers move together. Share of answers survives that, since a market-wide shift moves the denominator too. Name the competitors in the report, freeze that set alongside the prompt set, and use an identical denominator for every brand you compare. Our guide to running a competitive AI visibility analysis on share of answers covers the construction method, including the rivals an engine invents that your client has never heard of.
AI-referred sessions and conversions — Observed tier only. This is where GEO reports usually either overreach or lose their nerve. Aleyda Solis' three-layer framework splits business impact into four confidence tiers that should never be blended: Observed, Proxy–Own, Proxy–Third-Party, Modelled. Only Observed belongs on page one; modelled numbers like estimated influenced pipeline go later, labelled as modelled and expressed as a range. Her caveat is the sentence to steal for the footnote: "Measured AI referral traffic is the floor, not the ceiling, of AI's contribution."
Search Engine Land's executive-reporting guidance adds a warning: don't headline "AI-driven visits" as a standalone metric. AI referrals are a thin slice of traffic for most clients, so put the session count beside conversions or leave it off page one. The plumbing is fiddlier than it looks; connecting an AI visibility score to GA4 conversions covers what GA4 shows, what it drops, and how to recover the difference.
Prompts won and lost. The fourth number turns statistics into a work list: how many tracked prompts crossed into "consistently mentioned," how many fell out, and the names of both. It converts an abstract percentage into "we now show up for best CRM for field service teams, and we stopped showing up for HubSpot alternatives" — and it writes next month's scope for you.
This isn't the only defensible set. Similarweb's agency guide proposes Brand Visibility Score, Brand Mention Share, Domain Influence and AI Traffic — reasonable, with one substitution worth making. Move Domain Influence, the share of citations pointing at the client's own site, off page one: AirOps found brands are 6.5× more likely to be cited through third-party sources than their own domain, with roughly 85% of brand mentions originating off-domain. A number that falls when your off-site strategy is working won't survive the person signing the invoice.
How to present nondeterminism without eroding trust
Now the hard part: explaining, every month, in one paragraph, why the same question produces different answers — without making it sound like your measurement is broken.
The strongest available explanation isn't about AI being mysterious. It's about GPUs. Thinking Machines Lab generated 1,000 completions of the same prompt at temperature 0 from Qwen3-235B and got 80 unique completions, the most common appearing 78 times. All 1,000 were identical through token 102; at token 103, 992 said "Queens, New York" and 8 said "New York City." The cause was that inference kernels aren't batch-invariant — server load changes the batch size, batch size changes the floating-point arithmetic, and the arithmetic changes the token.
Translated for a client: the answer varies partly because of how many other people were using the model at that instant — a property of the infrastructure, not a signal about your brand. That's an honest reason their laptop disagrees with your dashboard, and one no vendor can engineer away, so you aren't implicitly promising to fix it next quarter.
The behavioural evidence is more reassuring than it sounds. AirOps found only 30% of brands stay visible from one answer to the next and just 20% persist across five consecutive runs — but roughly 50% of brands that disappear resurface within two runs. That last figure is the most useful thing you can hand a nervous client: it reframes "we vanished" as "we're in the rotation, not out of the set." Mike Sonders' experiment running 12 prompts 100 times each points the same way — a single prompt surfaced around 44 distinct brands across 100 responses, and only about 5 of them (11%) hit an 80%-or-better mention frequency. His conclusion travels well because it's third-party: "a single spot-check for any given prompt is nearly meaningless."
One more reason their check differs, worth a sentence in every report: memory, custom instructions, login state and location all change which brands get named, which is why practitioners now treat logged-out and logged-in as separate tracked surfaces. Your report measures a clean logged-out surface; your client's phone measures theirs. Both are real; they aren't the same instrument.
Four artefacts that make the honesty structural
Saying "AI is variable" once in an intro reads as excuse-making. What builds trust is putting the variance into the furniture of the report, so it's there in good months too:
- A confidence band on every trend line — not only the ones going down.
- Run counts in the footer: how many prompts, how many runs each, which engines, what window.
- A stated movement threshold on page one: "changes smaller than ±X pp are not treated as movement." Decline to claim a 4-point win and the client will believe you about a 4-point loss.
- A live re-run offer. When the screenshot email lands, the answer isn't theory — it's running the prompt set again and showing the distribution.
That last one is a tooling question. Geoptimizer's headline number is a 7-day rolling window with a confidence band rather than a single sample, single scans are labelled as snapshots, and a scan finishes in about 60–120 seconds — so you can re-run the client's prompt during the call and show ten answers instead of arguing with one. The formula is published rather than proprietary (EngineScore = 100 × (0.35·MentionRate + 0.25·CitationRate + 0.20·Prominence + 0.20·Sentiment), averaged across engines), so your methodology appendix can point at the evidence behind the visibility score formula rather than maintaining its own. Rand Fishkin told marketers to demand "stats-backed, publicly-reviewable research" from tracking vendors; a client asking you the same question deserves a URL, not reassurance.
The sample-size arithmetic you should be able to defend in a QBR
Confidence bands are computable, and an account lead should be able to produce the number under pressure. Using the standard proportion formula at 90% confidence with assumed visibility of 30%, Obsero calculates that ±10pp needs 57 prompt-readings, ±5pp needs 228, and ±1pp needs 5,682 — precision improving only with the square root of sample size. Their worked example: 100 prompts across three models, grouped into topics of 20 and refreshed daily, gives about 420 readings per topic per week (±3.7pp) and roughly 1,800 a month (±1.8pp).
Run that against a small retainer. Thirty prompts on four engines, one run each per week, is 120 readings weekly — scaling from the ±5pp/228-reading benchmark, roughly ±7pp for a single week, tightening to about ±3.5pp if you pool the whole month (and that pooled figure flatters you, since repeated readings of the same prompts aren't fully independent). So on that program a five-point week-on-week move is not a finding, it's the instrument. Put it in the footnote every month, and the month the number drops six points you'll have earned the right to say so.
Page two: per-engine, never blended
A composite belongs on page one because clients need one line to trend. The decomposition belongs on page two — not in an appendix, and not on request.
Profound tracked 15,155 brand-category pairs daily through May 2026 across three Google surfaces alone — the Gemini app, AI Overviews and AI Mode — and found a median 8-point daily visibility gap between each brand's best and worst Google model, with more than a third of brands showing gaps above 10 points. Position drifted about 0.53 ranks per day within a model, while the median gap across models was 3.36 ranks — roughly six times larger. The surfaces named similar quantities of companies per answer (4.4 to 5.0) but different ones, sharing only about a third between any two.
If one vendor's three surfaces disagree by 8 points, a blended cross-engine number hides exactly what the client pays you to find: which engine is failing them, and why. Aleyda Solis lists "blending multiple platforms into one score" among the named pitfalls of prompt-library measurement for the same reason. So: one composite headline with a published formula and a band, decomposed by engine on the very next page, every month.
When the score drops for reasons nobody controlled
Some months the line goes down and nothing in the retainer caused it. The last twelve months supplied a small library of examples, and having them dated in your appendix beats any amount of reassurance.
On 27 January 2026, Google made Gemini 3 the default model for AI Overviews. SE Ranking analysed 100,000 US keywords across 20 niches before and after and found roughly 42% of previously cited domains no longer appeared at all, while average sources per overview rose 31.8%, from 11.55 to 15.22. The client-report sentence writes itself: two in five previously cited domains were replaced across a 100,000-keyword sample. Your drop sits inside that event, not outside it.
Then, between 16 February and 2 March 2026, Seer Interactive — monitoring around 82,000 responses across 20 brand workspaces — watched Gemini's citation rate fall from 99% to 76%, a 23-point drop sustained through mid-March, with 11 of the 20 brands seeing shifts greater than 5pp. Listicle-heavy sources fell hardest, and "best of" prompts declined far more than "how-to" prompts. When a model leans on training data instead of live retrieval, recent editorial work stops showing up — so a client who spent the quarter on comparison pages sees a flat line they didn't earn.
Baseline churn does damage in quiet months too. SISTRIX's AI Research Index covered 82,619 qualified prompts and 1,548,213 snapshots across 17 weekly reference dates from December 2025 to April 2026, and measured weekly churn-in of cited domains at 74% for ChatGPT Search and 56% for Google AI Mode, against 5% for AI Overviews. In ChatGPT, the median prompt has no domain appearing in all 17 weeks. As the study puts it: "For 86% of all prompts, there is a stable core comprising a few domains; the rest rotates at a rate of 89% per week." Which is the best argument going for never sending a client a weekly citation list.
Sometimes the drop is in your measurement plumbing rather than the engines. ChatGPT's May 2026 homepage-link change pushed referrals to homepages from 26–29% of all referral traffic to 62–63% by late May, rearranging every client's landing-page mix overnight. GA4 added a native AI Assistant channel in May 2026 on the reserved ai-assistant medium, rolling out gradually and without reprocessing historical sessions — so a client running both a custom "AI Search" regex channel and the native one can see the same sessions counted differently in consecutive reports. And Search Console's generative AI performance report, documented in June 2026, is impressions-only: no clicks, no CTR, no query data, so it won't close the attribution gap either.
The triage order for any drop
Work the causes in this sequence, because the cheapest checks eliminate the most reports:
- Your own change log first. Did the prompt set, tool, engine mix or window change? Most common cause, most embarrassing to discover second.
- Measurement plumbing — GA4 channel definitions, referrer behaviour, tagging.
- Model or version change. Dated, industry-wide and easy to explain; our diagnostic for separating a real visibility shift from a model-version artefact is built for this step.
- Citation-mode change inside one engine, as in the Gemini February episode.
- Normal rotation — is the move inside your stated threshold and inside the band?
- Competitor gain. Share of answers is close to zero-sum; check who won what you lost.
- Genuine content or authority regression. Last, not first.
Then report the diagnosis, not just the number — and report it early. Search Engine Land puts the principle well: "If results are down, tell the client or the board early. Flagging a decline before you're asked about it preserves trust." An agency that opens the report with a Gemini-3-scale event looks like it's watching the category; the same agency answering that question three weeks later, by email, looks like it got caught.
The appendix nobody reads and every good report has
The appendix exists to be pointed at, not read. Four things belong in it.
Prompt-set version and change log. This is the most common way an agency accidentally misleads a client: add three prompts mid-quarter that the client is likely to win and the headline rises for free; drop a losing prompt and it rises again. Conductor recommends updating prompts on a predictable cadence such as quarterly or biannually, because ad-hoc changes introduce volatility and reset the baseline. Make it a rule with a version number: no changes mid-period, changes take effect at period boundaries, every change annotated on the chart.
Set construction, briefly. Aleyda Solis' prompt-library sizing guidance runs from 30–60 prompts for a single product and market up to 100–250 for SaaS with several personas, keeping a stable core while experimental prompts rotate. Conductor suggests 75% unbranded / 25% branded, and SE Ranking adds that branded prompts should be tracked separately rather than mixed in: with the brand name in the prompt, visibility is near-guaranteed and the blended number inflates. Aleyda's line is the one for the client — "Representative doesn't mean exhaustive."
Run counts, engines, model versions, window length, plus the tool name. Never change tools mid-engagement without restating the baseline in the same report.
Confidence tiers on every business number — Observed, Proxy–Own, Proxy–Third-Party, Modelled, labelled and never blended — plus a note on what you deliberately don't claim. Prompt-volume estimates are the obvious case: modelled from browser panels skewed toward tech-savvy users, they carry a "pseudo-accuracy" critique worth reading before a volume number goes in front of a CFO. Use them to prioritise, label them as estimates, keep them off page one.
Making it survive ten clients
A template that takes two hours per brand is one you'll abandon by month three. Three habits keep it inside the hour that 73% of agencies work within.
Automate the pull, not the interpretation: export visibility, per-engine and competitor data on a schedule, and spend the human time on the "what changed and what we did" page — the only part a client can't get anywhere else. Maintain one methodology page and link every client to it instead of pasting an appendix into twelve decks. And set expectations in the retainer scope before month one: monthly cadence, rolling four-week windows, a movement threshold, frozen prompt sets with quarterly revisions, and a first real review at 90 days.
One last honesty note, because it sets the tone for everything above. Being cited correlates with better click-through — Seer Interactive found cited brands saw organic CTR of 0.70% versus 0.52% uncited — but the study team declined to claim causation, noting it's equally possible that stronger brands are simply more likely to be cited. Reproduce that caveat when you use the number. Clients remember the agency that told them what the data couldn't prove.
FAQ
How often should we send an AI visibility report to clients? Monthly, matching the 69% of agencies already on that cadence. Weekly reads are for internal anomaly detection — with weekly citation churn at 56% on Google AI Mode and 74% on ChatGPT Search, a weekly client report is a subscription to explaining noise.
What do I say when a client runs the prompt and doesn't see their brand? That one answer is one draw from a distribution — then show the distribution. Two runs of the same prompt return the same brand list less than 1% of the time, and personalisation and login state shift it further. Re-run the set live and point at the rate rather than the screenshot.
How many prompts do we need before the number means anything? Roughly 57 prompt-readings gets you ±10pp at 90% confidence; 228 gets you ±5pp. In practice that's 30–60 core prompts run three to five times per engine per week, pooled over a rolling four weeks. Below that, report a wide band and a movement threshold, and don't claim wins inside it.
Can we put AI-influenced revenue on page one? Put Observed conversions on page one and modelled influence later, as a range and labelled as modelled. Measured AI referral traffic is the floor of AI's contribution rather than the ceiling — but an unlabelled modelled number is one CFO question away from costing you the account.
Should we change the prompt set when the client launches something new? Yes, at the next period boundary, with a version bump and an annotation on the chart. Mid-period edits reset the baseline, and they look indistinguishable from cooking the number even when they aren't.
The short version
Four numbers on page one, a per-engine decomposition on page two, a diagnosis rather than a number when the line moves, and an appendix stating the prompt-set version, run counts and confidence tiers. What makes it work isn't the layout — it's declining, month after month, to claim a win you can't distinguish from noise, so your explanation of a loss carries weight.
If you'd rather the measurement layer made that argument for you, Geoptimizer scores ChatGPT, Gemini, Claude and Grok separately before averaging them, publishes the formula and its version history, and reports a rolling window with a confidence band instead of a single sample. The 14-day free trial includes your first visibility report — see the current plans and what's coming for multi-brand teams before scoping your next GEO retainer.