Short answer, as of August 2026: don't sell a new service line yet — sell a two-week baseline. Charge for a diagnostic that establishes where the client stands across the major engines, then scope the actual work from what you find. Most clients turn out to need mechanical fixes worth a few hours, not a retainer. The ones with a genuine content gap will be obvious, and you'll be quoting from evidence instead of hope.
Here's the 30 days, week by week, including what to say when the client asks for a rank number and there isn't one.
Week 0: the conversation before the plan
Your client asked "how do we rank in AI search." Three things need saying before any work starts, and saying them early is what keeps the engagement out of trouble later.
There are no rankings. There is inclusion or absence in a composed answer. Position 7 doesn't exist here — you're named or you aren't. That single reframe changes the whole reporting conversation, and it lands far better in week 0 than in month 3.
Answers are non-deterministic. The same prompt asked twice gives different sources. Any single screenshot — theirs showing a competitor, or yours showing a win — is anecdote. This is the one clients find hardest, and it's worth spending real time on because every subsequent report depends on them accepting it.
Citations often produce no referral traffic. Someone asks an assistant, gets named your client, then calls them or searches the brand. Analytics records a direct visit. If you promise measurable referral traffic from AI citations you will be explaining a shortfall for the rest of the relationship.
Set those three expectations and you can run an honest program. Skip them and you're building on a promise you can't keep.
Week 1: build the prompt set and baseline
The prompt set is the whole measurement apparatus. Get it wrong and everything downstream is noise.
Write prompts, not keywords. Nobody types "commercial roofer dallas" into an assistant. They type "who should I call about a flat roof leak on a commercial building in Dallas, and roughly what does that cost." Interview the client's sales team for the questions prospects actually ask — that's the source material, and it's free.
Aim for 15 to 30 prompts spanning four intents:
- Direct recommendation — "recommend a [category] in [market]"
- Problem-first — the symptom the customer has before they know what to buy
- Comparison — "[option A] vs [option B]," where the client's category is one option
- Cost — "how much does [service] cost," because these get asked constantly and are often answered badly
Then sample. Each prompt, at least three times, across every engine you intend to report on — ChatGPT, Gemini, Perplexity, Claude, Grok, Microsoft Copilot. That's 15 prompts × 6 engines × 3 repetitions = 270 checks for one client. Do it manually once so you understand the texture of the data. Do not do it manually twice; that's what scheduled sampling is for, and it's the difference between a service you can run for forty clients and one you abandon by month two.
Record for each result: which engines named the client, which named competitors, which named nobody relevant, and what source URLs were cited. That last column is the most actionable thing in the entire dataset — it tells you exactly which pages the engines consider authoritative in your client's category.
Our writeup on how AI visibility tracking works covers sampling design in more depth if you want the reasoning behind the repetition counts.
Week 2: the mechanical audit
While the baseline is running, audit. Four checks, in this order, because each one makes the next meaningful.
Check 1 — AI crawler access. robots.txt for GPTBot and OAI-SearchBot (separate agents, per OpenAI's bot documentation), PerplexityBot, ClaudeBot, Google-Extended, Applebot-Extended, CCBot. Then the layer robots.txt doesn't cover: request the homepage with a non-browser user agent and confirm real HTML comes back rather than a 403 or a bot challenge. A CDN bot rule blocks crawlers regardless of what robots.txt permits.
Check 2 — rendering. Read the raw HTML response body, not the browser's rendered DOM. If services, locations, pricing or article substance only appear after JavaScript, fetch-only crawlers see a shell. Bing's webmaster guidelines have recommended initial-HTML delivery of critical content for years and it remains the single most consequential technical requirement here.
Check 3 — entity consistency. Name, address, phone, categories, service list and the one-line description, compared across the site, its structured data, Google Business Profile, Bing Places, Apple Business Connect and the top directories in the vertical. Models corroborate before they assert; contradiction reads as low confidence and low confidence names someone else.
Check 4 — structured data. Present, valid, and generated from the same record the page renders from. Schema.org defines the vocabulary; the discipline is that hand-authored per-page markup drifts from visible content the first time anything changes, and markup contradicting the page is worse than none.
In roughly half of first engagements, checks 1 and 2 explain the entire problem. That is a genuinely good outcome — it's a cheap fix with a large effect, and it's a fast, credible win to open a relationship with.
Week 3: fix, then prioritise content
Ship the mechanical fixes first. They're cheap, they're fast, and no content work matters until a crawler can fetch the page.
Then use the baseline's citation column to prioritise. You now know which pages the engines already trust in this category — usually a mix of competitors, trade publications and directories. Two questions follow:
Which of the client's existing pages are close? Pages that already rank conventionally are already retrievable. Upgrading one to the citation standard is a small edit to a working asset. Writing a new page means earning retrieval from scratch. Start with the former, always.
What does the citation standard require? The Princeton-led GEO study tested this directly and found that adding citations, statistics and quotations to source content lifted visibility in generative engines by up to 40%, while keyword stuffing produced nothing. Operationally, on each priority page:
- Replace vague quality claims with specific verifiable ones
- Add real numbers — ranges, durations, counts — wherever the client actually knows them
- Cite genuine external authorities for factual claims about standards or research
- Answer the literal question in the first two sentences of each section
- State the exceptions, because exceptions are what make a source read as expert
The bottleneck is always the same: the specifics live in the client's operation, not in your writers' heads. Build a subject-matter extraction step — a 45-minute call with the person who does the work, recorded, transcribed — and your output quality changes immediately. Our post on scoring a draft for AI visibility covers how to check a page against this standard before it ships.
Week 4: reporting the client will accept
This is where most agency AI-search programs die, because the report has no rank number and looks thin next to a conventional SEO deck.
Report these four things:
| Metric | What it is | Why it's defensible |
|---|---|---|
| Share of voice | % of sampled responses naming the client | Isolates the variable; comparable over time |
| Engine coverage | Which of the 6 engines name them | Shows where the gap is concentrated |
| Cited sources | Which URLs engines quote in this category | Directly actionable target list |
| Prompt-level gaps | Which specific prompts never name them | Becomes next month's content brief |
Show a trend line, not a snapshot. Non-determinism means any single reading is noise; the whole value is in the direction over eight or twelve weeks. Say so on the slide, every month, until it's internalised.
And support it with two conventional numbers that tend to move alongside AI visibility: branded search volume in Search Console, and self-reported attribution from the client's intake form. Adding "how did you hear about us?" to a contact form is embarrassingly low-tech and remains one of the better instruments available for this.
What to charge
Three honest structures:
The diagnostic (weeks 1–2), fixed fee. Baseline plus mechanical audit plus a written diagnosis. This is the product to lead with, because it's small, self-contained, produces a real deliverable, and scopes everything that follows from evidence.
Retained monitoring, monthly. Recurring sampling and reporting. Price it against your tooling cost — this should be a healthy-margin line because the marginal cost per client is near zero once the sampling is automated.
Content and fixes, scoped per finding. Quote from the diagnostic, not from a template. Some clients need four hours of engineering. Some need a quarter of content work. Guessing before the baseline is how agencies end up delivering the wrong thing at a fixed price.
The operational trap
At one client, running 270 manual checks a month is annoying. At fifteen clients it's over 4,000, and the program will quietly stop happening around week six — not because anyone decided to stop, but because it's the task that slips when a client emergency lands.
That's the argument for automating the sampling half and keeping the judgment half human. DigiRank's AI Visibility Tracker runs prompt sets across all six engines on a recurring schedule with multi-tenant separation and white-label share links, which is the shape agencies need; the Agency plan is $249/mo and covers up to 15 client tenants with a 14-day trial, and it connects to Search Console, GA4 and Google Business Profile so the AI data and the conventional data land in one report.
What stays human: choosing the prompts, interviewing the client's technicians for specifics, deciding which pages deserve the upgrade, and explaining non-determinism to a sceptical marketing director. No tool does those, and any vendor claiming otherwise is selling generated filler. We drew the same line for local work in what local SEO automation genuinely automates, and it holds here.
If you're pitching this next week
Lead with the diagnostic. Set the three expectations from week 0 in the first meeting. Show the client a real sample of what a baseline looks like — even from a demo account — because the citation column tends to sell the engagement by itself when they see which sources the engines currently trust instead of them.
And if the client's real question is "should we panic," the useful answer is no: most of the work that makes a business visible to AI engines is work a competent SEO program was already doing. Our breakdown of GEO versus traditional SEO budgets is a good thing to send them before the scoping call.
Frequently asked questions
My client asked how to rank in AI search and I don't know where to start — what's the first step? Sell a two-week baseline rather than a service line. Build a set of 15 to 30 real buyer prompts, sample each at least three times across ChatGPT, Gemini, Perplexity, Claude, Grok and Copilot, and run a four-point mechanical audit. Scope the actual work from what that finds.
How many prompts should an AI visibility baseline include? Fifteen to thirty, covering four intents: direct recommendation, problem-first, comparison, and cost. Source them from the client's sales team — the questions prospects actually ask — rather than from a keyword tool, because people speak to assistants in sentences.
How do I report on AI search when there are no rankings? Report share of voice across the sampled prompt set, engine-by-engine coverage, the source URLs engines cite in the category, and prompt-level gaps. Always show a trend over eight to twelve weeks rather than a snapshot, because individual readings are noise on a non-deterministic surface.
What should I charge for AI search work? Lead with a fixed-fee diagnostic covering the baseline and audit, then quote content and fixes from the findings. Retained monitoring should be a separate monthly line priced against your tooling cost, where marginal cost per client is near zero once sampling is automated.
How much of an agency AI search program can be automated? The repetitive half: prompt sampling across engines, crawler and rendering checks, entity audits, scheduled reporting. The judgment half cannot be — choosing prompts, extracting specifics from the client's technicians, deciding which pages to upgrade, and setting expectations with stakeholders.
Do most clients actually need a big content project? No. In roughly half of first engagements the entire problem is mechanical — blocked AI crawlers or client-side-only rendering — and the fix is a few hours of engineering. Running the diagnostic first is what stops you selling a content retainer to a client who needed a robots.txt edit.
How long before a client sees movement? Mechanical fixes surface within days to weeks once pages are re-fetched. Content upgrades on already-ranking pages take weeks. Entity consistency propagates slowly. Off-site corroboration takes quarters. Set that timeline in the first meeting, not in the third monthly report.