How to review a quarter of AI search work with a client

A quarterly AI search review with a client should separate four things: the work you completed, what you observed in AI answers, what is still unresolved, and the options for next quarter. Build the deck from saved records, not memory: a weekly report for the first and last completed week, an export of the quarter's data, and the status of every content fix. Present observed changes as observations, never as proof your work caused them. Then let the client choose between named options, so the meeting ends with a decision rather than a recap.

In short

• Structure the review as four parts: work done, changes observed, unresolved issues, next-quarter options.
• Compare the first and last completed weeks of the quarter, and name every change in prompts, engines or scope between them.
• Report status for every fix, including the ones that were dismissed or went stale, so the deck does not read as a highlight reel.
• Say plainly that an outcome check is an observation, not a proven cause.
• End with two or three options the client can pick from, each with what it would and would not change.

What should a quarterly AI search review actually cover?

A quarterly review answers four questions in order: what did we do, what did we see, what is still open, and what should we do next. Anything else is decoration.

A weekly review is an operating meeting. The quarterly review is different. Its audience is the person who pays for the work and decides whether to continue it, and they are asking whether the effort was sensible, not what moved on Tuesday.

That has one consequence for how you build it. A weekly meeting can be informal because the next one corrects any error. A quarterly review is the record the client will remember, so every number in it needs a source you can point to.

What evidence should you gather before the meeting?

Gather five things: the first and last weekly reports of the quarter, a data export for the full window, the status of every content fix, the list of open items, and a note of every scope change. Do this a few days ahead so gaps surface before the client sees them.

Here is how each maps to the product.

Weekly reports for the start and end of the quarter. The weekly report is built for the last complete Monday to Sunday week in UTC. In Client reports you can choose a common report week and open the exact report for each project. A card reads "No generated report for this week" if none exists, and the product never substitutes a different week, so check both endpoints exist early. Nothing in the app or the admin panel builds a report outside the weekly run, so a missing week cannot be generated after the fact. The report's numbers come from the same figures the product already calculates, with no language model in the loop, per the weekly report documentation.

An export for the whole quarter. Exports offers thirteen datasets as CSV or JSON. Twelve are windowed by the dates you choose, so set the window to the quarter explicitly. Each dataset has a row cap and a download that would exceed it is refused rather than truncated, so narrow the window or download in parts. Answers is capped more tightly than the rest because each row holds a whole response. Read CSV files by header name, not column position, because column order can change.

Fix status. The Optimizations screen holds each drafted fix with a status: ready, applied, stale or dismissed. The Page issue queue shows citation gaps and optimizations grouped by page, but it covers known pages only, not a full site crawl, and it does not list every historical recommendation. Treat Optimizations as the record of fixes.

Open items. The Actions To do list gathers open records from seven parts of the product, grouped into High, Medium and Low bands. It has no status of its own; each item is a record living elsewhere. Snapshot it before the meeting, because it lists only what is open today.

Scope changes. This one is yours to keep. The Prompts export is the project's current configuration, not a history, and the docs describe nothing that records when you added a competitor in week 5 or paused six prompts in week 9. Keep a change log as you go, including any engine that began or stopped answering.

Enlarge image

Enlarge image

How do you compare the start and end of the quarter fairly?

Compare like with like: the same prompts, the same engines and the same definition of each metric, and state any way in which they differ. A before and after comparison is only as honest as its list of differences.

Start with which metric you are quoting and what it divides by. Brand visibility is analysed answers naming your brand divided by analysed answers. Citation rate is collected answers that cited your domain divided by collected answers. Their denominators differ, so the two can legitimately disagree on the same screen (the metrics page walks through an example). Put the definition in a footnote on the slide; the client will otherwise assume they are the same thing.

Then check three sources of drift:

• Prompt set. If you added or paused prompts, the two weeks measure different questions. Say which.
• Engines. A weekly report compares a week with the week before it, and where an engine was not measured in both weeks it compares only the engines measured in both and says so in its summary. That covers adjacent weeks only. Your first-to-last comparison spans a quarter, so check yourself whether any engine started or stopped answering in between, and say so on the slide.
• Sample size. Below 30 observations a number is provisional. If a per-prompt figure rests on a handful of answers, say so or leave it out.

Related reading: Why your AI visibility score changed when your website did not and Did your content update help? A practical before-and-after review.

What goes in the deck? A 12-slide outline

Twelve slides is enough: two to frame, seven to report, one to be honest about gaps, and two to decide. Copy this table into your deck tool and fill each row from the evidence above.

Slides 4 and 5 should sit side by side, with the same visual weight. A deck that shows only completed work invites the client to ask what you left out.

Enlarge image

How do you report what happened without claiming it worked?

Say what you observed, what you did, and that the two coincide in time; do not say one produced the other. The product's own documentation takes the same line.

The optimization outcome check illustrates the limits. After you mark a fix applied, a later check compares visibility for the fix's prompts against the 30 days before you marked it applied and returns a label: "positive", "negative", "neutral" or "pending more data". At the month mark, positive needs five points or more, negative needs five points or more the other way, and a smaller change reads neutral, which means nothing definitive happened yet, not necessarily that the fix failed. Under a week, it always reads "pending more data"; between a week and a month it stays pending unless visibility has already risen ten points or more. The Page issue queue documentation says the same of an applied date and outcome: saved observations that do not prove the edit caused a change.

So the slide wording matters. Write "Visibility on the 4 affected prompts rose 6 points in the month after the edit", not "the edit raised visibility by 6 points". Several things move at once in AI answers: models change, competitors publish, and your prompt set may shift. Lay out other candidate explanations on the slide rather than waiting for the client to raise them.

A citation, a mention and a stated reason are all observations of an answer. None of them reveals why a model produced the answer, so avoid sentences that begin "AI now prefers us because".

What should you do about the work that did not go well?

Show it, count it and give each item a next step. Dismissed drafts, stale drafts and fixes still pending measurement are part of the record, and hiding them undermines the numbers you do show.

Three statuses deserve a sentence each. A dismissed draft was a choice: someone decided against it, and the deck should say who and why. A stale draft was written against a page that has since changed, so it needs a fresh draft before anyone acts on it. An applied fix in "pending more data" cannot be judged yet; say when the check will be meaningful.

Also list what the platform cannot tell you. Optimizations never puts a draft on the client's site; someone makes the edit and marks the fix applied, so a fix counts as applied only when that happened. If the client's team did the edit, say so; if a fix sat waiting on their approval, put that on slide 10 as a blocker with an owner.

What options should you offer for next quarter?

Offer two or three options, each defined by what it would change and what it would leave alone. This turns the meeting from a report into a decision.

Draw the options from the evidence rather than from a template:

• Deepen: more fixes on pages already in the queue, where gaps remain open.
• Widen: new prompts, markets or competitors, accepting that a wider set changes what later comparisons mean.
• Hold and monitor: keep collection and weekly reports running, with no new content work.
• Change approach: for example, an earned mention or a new article instead of another page edit, where the evidence points that way.

If you are scoping the next quarter formally, see Scope an AI visibility engagement around deliverables you can control. Do not promise a level of visibility in any option; promise the deliverables you control.

Worked example: Northfold Digital reviews Quillstone's quarter

Northfold Digital monitors Quillstone, a document-review software company for legal and compliance teams, against fictional competitors Brieflane and Clausewise. For the Q3 review, Northfold pulls the weekly reports for the first and last completed weeks and a quarter export.

Before presenting, Northfold checks its change log. In week 6 it added five prompts about compliance workflows, and in week 9 one engine started answering. Northfold puts both facts in the slide 6 footnote and adds a line: "The last-week figures include five prompts added in week 6 and an engine that began answering in week 9." Northfold also offers to restate both weeks on the original prompts and engines only, working from the Answers export outside the product.

On work: eight drafts were reviewed. Five were applied, two were dismissed because the client's legal team objected to the wording, and one went stale when the client redesigned that page.

Slide 9 reads: "Two of five applied fixes show a positive outcome label; one is neutral; two are too recent to judge. These are observations on the affected prompts, not proof of cause." Slide 10 lists the stale draft and the two pending checks, with a date for each. Slide 11 offers three options: deepen on the open queue, widen to a second market, or hold and monitor. Slide 12 asks the client to choose one and name the legal contact for future wording reviews.

Notice what the deck does not say: that Quillstone "won" the quarter. It shows movement, its caveats, and a decision.

Common mistakes and limits

• Comparing different prompt sets. If prompts changed, the start and end weeks are not the same measurement. Disclose or restrict to the prompts present throughout.
• Choosing the best week. Pick the first and last completed weeks in advance, not the strongest ones.
• Mixing denominators. Brand visibility and citation rate divide by different populations; do not place them in one sentence as if they were one measure.
• Treating "neutral" as failure or "positive" as proof. The label is a threshold on a change in visibility, nothing more.
• Presenting a share link as a live dashboard. A client digest link is fixed at creation and later report regeneration does not update it. Create a new link for a newer copy. See Client report links.
• Ranking clients against each other. If you hold several projects, the Client reports overview is explicit that its cards are separate readings, not an agency-wide average. Prompts, countries and engines can differ between clients, so do not turn them into a league table.

What this review cannot tell you: why a model gave any particular answer, whether a change will persist, or what would have happened without your work.

Frequently asked questions

How long should a quarterly AI search review take?

Plan for 45 to 60 minutes with the decision-makers: about 15 minutes on work and observations, 15 on unresolved issues, and the rest on options and decisions. Send the deck ahead so the meeting is spent deciding, not reading.

Should I share the live app or a report link with the client?

That depends on what you want them to see. A client link is a fixed digest of one weekly report, opened without an account, and you choose its expiry of 1, 7 or 30 days. Anyone holding the link can open it, so treat it as a credential. Check the sharing documentation for exactly what the digest includes and excludes.

What if the numbers went down this quarter?

Show them with the same structure. Check first whether the prompt set or engines changed, then look at which prompts fell and what changed on the cited sources. A decline with a clear list of what you checked and what you propose is more credible than a vague explanation.

Can I combine several clients in one review?

A combined agency report covers 2 to 10 ready client projects, needs owner or editor access to each, and keeps each client's metrics separate with no pooled average. For a review with one client, use that client's own report and export. Do not bring another client's data into the room.

Is a quarter the right interval?

It is a common one, and it gives enough time for a page edit to be checked over a full month. But if you agreed a different cadence or the engagement started mid-quarter, review on your agreed dates and use the same structure.

Next step

Pull the first and last weekly reports of your quarter, download the quarter export, and fill the 12-slide table above with real records before you write a word of narrative. If you want the reports, exports and fix tracking in one place, sign in to DiscoveredBy.