A visibility score and a promise to get recommended in ChatGPT are not a case for $4,000 a month. If you need ongoing help across Google, ChatGPT, Gemini and Perplexity, but have to justify it to a skeptical founder or finance lead, bring evidence instead. I used to think a nice chart could carry that decision. It cannot. These seven exhibits define what you are buying, who will deliver it and how you will judge it, without pretending rankings or AI recommendations come with guarantees.
1. Bring a buyer-question map, not a keyword-volume export
What finance hears when you say “keywords” is traffic. What I want them to hear is the set of questions that decides who makes the shortlist.
A keyword-volume export lists 500 phrases sorted by searches per month. A buyer-question map lists the 20 to 30 questions that actually influence who gets bought:
- Best tool for [use case].
- [Your brand] vs [competitor].
- How much does [category] cost?
- What breaks during implementation?
For each question, log which engines matter and what their answers cite today. That table defines the retainer’s scope. Without it, you are buying “SEO and GEO coverage” with no definition of covered.
Ask any provider pitching ongoing SEO and generative engine optimisation services to run that set several times per engine over two to four weeks before setting alerts. Log brands, order and cited sources each time.
The mechanism is simple: one answer tells you little. Thirty prompts across three engines, repeated over time, show where you lose the shortlist and where content or third-party proof could help. The map gives the work a commercial purpose before anyone starts counting articles.
First step: Pull your last 15 lost deals and write down the exact question each buyer asked before choosing. If those questions are missing from the map, fix the map.
No map, no monthly fee.
2. Bring a reproducible baseline across Google and AI answers
One screenshot of ChatGPT naming you proves nothing about consistent visibility. Answers change between runs and across days. A Sistrix analysis of 10,826 daily series found 70% of brands appear on fewer than 1 in 5 days, and only 6% appear on 4 or more of 5 days across ChatGPT, AI Overviews, AI Mode and Perplexity.
A Tuesday mention is not a Thursday promise.
Ask for a baseline you can rerun to understand the pattern, not reproduce an identical answer. It needs:
- A fixed prompt list.
- Named engines and recorded dates.
- Several runs per prompt.
- A log of brands and cited sources.
- The variation, not just the best-looking result.
The math rewards breadth over repetition. Twenty different prompts measured once gives you about plus or minus 9.5 points, versus plus or minus 34 points for one prompt measured over 20 days. Around 100 prompts brings that to plus or minus 4.2 points. A starting set of 25 buyer questions gives you a defined scope; it does not make a precise market-wide visibility estimate.
Even citations move around. Same-day reruns repeat cited domains only 41% of the time in ChatGPT and 45% in Gemini, versus 94% in Perplexity. That is why a ChatGPT screenshot is not a shortlist.
First step: Freeze your 25 prompts and run them three times this week. Save the dates, brands and citations. If a vendor will not show variation, they are selling the flattering run.
3. Bring an access and indexing audit
Before paying for new pages, prove Google can reach the old ones. Important pages can be blocked, left noindexed after a redesign or never submitted because nobody owns the property. That is a discoverability problem, not a writing problem. Buying articles will not fix it.
Bring an audit that separates those problems, with read access only at this stage. Google’s own hiring guidance recommends read access to Search Console for the audit, not write access. The audit should give realistic estimates of the work involved, not promises.
If a vendor demands admin access on day one just to tell you what is wrong, walk away.
The ownership questions come first for a reason. Put these properties in one table:
- Search Console.
- Analytics.
- Business Profile.
- Hosting.
- Content and the CMS.
Record who owns each property, who has a named login and what access the proposed work requires. I want that sorted before a contract turns an access problem into a monthly invoice.
First step: Build the ownership table. If your company does not own its own Search Console property, fix that before signing anything.
No crawl access, no content budget.
4. Bring a publishing plan tied to gaps buyers actually encounter
A proposal listing 12 blog titles and a word count does not explain what the business is buying. Finance should reject it.
Ask for one row per proposed page, with the buyer question it answers, the evidence it will contain and why it deserves to exist. More posts without missing proof just means more pages for Google to ignore and for ChatGPT to skip in favor of a third-party review that actually names prices, limits and comparisons.
The gap should determine the page. If your baseline shows you lose on best tool for X because answers cite a competitor comparison you do not have, the plan should propose that comparison. Specs, screenshots and pricing logic, not another thought-leadership post. If you disappear on implementation questions, propose a setup guide, not a glossary.
A useful way to frame the work is to diagnose the four gaps before writing more content:
- Missing page: The buyer’s question has no useful destination.
- Thin page: A page exists but lacks the proof the question needs.
- Uncrawlable page: The content exists but cannot be reached.
- Uncorroborated claim: Your claim lacks supporting third-party evidence.
Each gap has a different fix. Only two primarily involve writing pages. A publishing calendar that treats all four as “create more content” is selling production before diagnosis.
First step: Take your 10 highest-value losing prompts and ask the provider to map each to keep, rewrite or create, with an owner and date. Record any access or corroboration work separately rather than disguising it as a new article.
If a title cannot name the buyer question it serves, cut it.
5. Bring a technical execution owner
Recommendations are cheap. Someone has to touch the CMS.
The stalled-retainer pattern is familiar: the agency delivers a 40-page audit, your developer never has time, and month three pays for a PDF. Bring a one-page ownership sheet before asking for money. It needs to answer three questions:
- Who can publish without further approval?
- Who approves the remaining work within 48 hours?
- What needs your developer, and how many hours per month will it take?
If the provider cannot answer, you are buying advice without a delivery plan. That may produce a handsome document. It does not put the fix live.
This separates tools that report issues from services that ship the fix: who logs into what, changes what and shows proof it went live. groas runs as a done-for-you service here, with a named strategist owning the result and changes accompanied by reasoning. That is the model I want: execution with accountability, not another queue of recommendations for the client.
First step: Ask the vendor to mark every audit item as provider, client or developer, with a date. Then check whether the people assigned actually have the time and access.
If everything says “client to implement,” you are not buying done-for-you execution. Do not pay as though you are.

6. Bring an authority plan with named sources, not a citation count
A vendor managing how your brand appears across ChatGPT, Gemini and Perplexity can promise 500 citations by next quarter. The number means nothing until they name where those mentions are supposed to come from.
Ask for 10 to 15 real places: the trade publications your buyers trust, the review and comparison sites AI answers already quote in your category, partner blogs, local press or community threads they can actually pitch.
The plan needs three columns:
- Source: The actual publication, site or community.
- Reason: Why it matters to buyers and the answers you are tracking.
- Owner: Who handles the outreach.
The mechanism is why finance should care. A Muck Rack study of 40M+ AI answers and 207M citations found that 84% or more of citations are earned third-party mentions, 15.2% are brands’ own sites and 0.3% are paid. Your blog does not vote for you. Other sites do.
That makes outside corroboration part of the work, not a vague bonus after the content calendar. It also makes a named source list more useful than a large citation target. You can inspect the former. The latter fits nicely in a sales deck.
First step: Ask for five placements the provider earned in the last 90 days, with live URLs, for a business like yours. If they cannot show the last five, do not accept a promise of the next fifty.
Unnamed citations are rumors.

7. Bring a report that separates shipped work, visibility and revenue
Finance does not need another dashboard. It needs three layers kept apart:
- Work: What was completed, verified or blocked, with an owner on each line.
- Visibility: Search Console clicks, impressions, CTR and position, alongside AI mentions and citations.
- Business outcomes: Qualified enquiries, orders or attributable revenue, with definitions attached.
That separation is central to useful monthly SEO reporting: publication is not proof of impact. I will take a boring table that respects the distinction over a glossy visibility score every time.
An AI mention is exposure, not a sale. Answers can cite you for two weeks while pipeline stays flat because they recommend a competitor alongside you, or because the cited page has no quote form. Mentions and citations belong in the middle layer. They help diagnose visibility; they do not automatically establish attributable revenue.
Apply the guarantee test, too. Google says no one can guarantee a number-one ranking and warns against providers guaranteeing first place or claiming a special relationship. Swapping “first place” for “guaranteed ChatGPT recommendation” does not make the promise credible.
First step: Ask for one sample report before signing. If work, visibility and revenue are blended into one score, you will not know which part of the service worked. No separation, no renewal.
Then turn the seven exhibits into a one-page approval memo. Finance will not read seven PDFs. Give it five lines, backed by the evidence:
- Scope: Which 25 prompts and engines the fee covers, using the map and baseline.
- Internal workload: Who approves within 48 hours, who owns access and how many developer hours the work requires each month.
- Delivery: What counts as completed work, who ships it and how it is verified.
- Fee and exit: The fixed monthly amount and how you cancel if visibility never moves.
- Uncertainty: The variation you actually measured, with visibility explicitly separated from revenue.
That memo turns “trust us on AI search” into a purchase with boundaries. If a provider will not put their name under those five lines, you have your answer before money moves.
If I had to start with one exhibit, I would start with number two. The baseline gives you something concrete to challenge before you buy. Run your 25 prompts three times this week, save dates and citations, and watch which vendors flinch when you show variation. The ones who explain it understand how your brand shows up across engines. The ones relying on a single glowing screenshot have not made their case. Start there, and make the other six earn their place.

