Do not buy the 1,000-prompt AI visibility package. A large prompt library can make your report look comprehensive while missing the questions that decide pipeline.
Leadership wants a share-of-voice number for ChatGPT and Claude. Vendors are happy to sell one. But coverage is not a plan. Tracking more prompts does not establish that any buyer asked them, much less that the answers matter to your business.
I start in the opposite place: search terms that produced real conversions. I turn those into natural questions, then use customer language and competitor citations to expand the watchlist. The list starts smaller. Every question has to earn its place.
A thousand prompts can hide the questions that matter
The measurement explains the miss. You give a tracker 500 or 1,000 suggested prompts, it runs them across ChatGPT, Claude and Gemini, then rolls your appearances into an AI share-of-voice number: your brand mentions divided by total brand mentions for the fixed prompt set, measured separately per assistant.
Change the set and the score changes. That makes the choice of questions part of the result, not a bit of setup to rush through.
One domain’s February 2026 ChatGPT report made the problem plain: recommended in 14 of 20 category prompts but absent from all 20 tool-intent prompts where buyers build a shortlist. The category coverage looked healthy. The shortlist coverage was zero.
Practitioners who track this for a living make the same distinction: tracked prompts should represent questions people ask when researching, comparing or choosing a provider. Think best alternatives to HubSpot, not just what is a CRM.
Revenue sits close to a decision: comparisons, alternatives, pricing and implementation. An informational library can run for months without touching those decisions. Before buying more coverage, I want to know what the current list leaves out.
Ask Claude for the searches that earned their place
I run this inside Claude with groas connected over MCP. The first prompt is boring on purpose:
Using our connected ad accounts, identify search terms that produced conversions. Show the conversion action, spend, date range and data limitations. Make no changes.
That last sentence matters. Evidence first. No edits, no new campaigns.
Paid search provides one of the cleanest environments for identifying queries used to reach your site: someone typed a search, saw your ad, clicked and did something you paid to track. With no prompt-volume data in AI search, a converting search term gives me a stronger starting point than a clever prompt idea. It does not prove someone asked an assistant the same question. It proves the underlying need deserves attention.
Then I interrogate the receipt. A form fill is not automatically qualified pipeline. Google Ads optimizes toward whatever conversion is marked primary. If that is a raw form fill, it can find cheap form fills, junk included. In one 2026 B2B SaaS audit, 78% of form fills never became an SQL.
I ask Claude to separate terms by the downstream outcomes available: MQL, SQL or closed-won, connected through offline conversion import tied to the original click. If the account only records fills, that limitation belongs beside the findings, not in a footnote nobody reads.
Do not rank buyer questions by unqualified volume and call it commercial evidence.
Translate buyer decisions, not just keywords
My next prompt keeps the distinction between evidence and inference visible:
Group these converting searches by buyer decision. Suggest natural-language questions for each group, and separate observed language from your inferred prompts.
A search term like hubspot alternatives pricing might become: What are the best alternatives to HubSpot for a 20-person team, and what do they cost to implement?
That is a useful candidate. It is not an observed customer quote. The team size and implementation detail came from Claude unless another source supports them. Keep the observed words separate from the invented phrasing. Otherwise, 40 polished sentences become “demand” through the administrative miracle of entering a spreadsheet.
Decision prompts include best-tool, head-to-head and alternatives questions. I group the searches around the choice underneath them:
- Compare: Which option fits the need?
- Replace: What should replace the current tool?
- Price: What will it cost?
- Deploy: What does implementation involve?
The ad account is evidence, not the whole market. It misses questions you never bid on and objections sales hears after the click. I expand the list with support tickets, sales transcripts and Reddit threads for customer language unedited by marketers, plus GSC question queries with impressions and competitor ad activity. With groas connected, Claude can read Reddit and competitor ad coverage alongside the search terms.
My admission rule is simple: a question needs a source. A converting term, documented customer question or recurring objection earns consideration. A fluent sentence from a prompt generator does not. Twenty buyer questions beat 500 guesses.
Lock the watchlist before measuring who wins
My third prompt prevents the report from quietly changing its own exam questions:
For this approved watchlist, show our mentions, competitor mentions and cited sources by assistant. Keep the prompt set and collection conditions explicit.
I use fixed questions, a fixed competitor set and same-day collection, reported separately by assistant. Share of voice is still just your mentions divided by total mentions for that set. A comparison needs a stable question set and stated collection conditions.
For the mechanics, run this 3-pass check for whether ChatGPT recommends your business, then bring the result back to Claude through groas and compare it with the paid-search evidence.
I sort the misses into three buckets:
- Coverage gap: The brand is absent while a competitor is named with a citation.
- Recommendation gap: The brand appears, but the answer recommends a competitor. It may even cite your page while recommending someone else.
- Generic answer: The answer cites nobody and gives general advice.
Those are different jobs. Bucket one points to missing coverage. Bucket two calls for examining why the answer favors another option; the relevant source might be your page or a third-party page you do not control. Bucket three is not automatically a defect to fix.
A mention means the brand appears in the text. A citation means a page is linked as a source. Track them separately or celebrate links that sent the shortlist elsewhere.
Competitor citations also help expand the watchlist. I read them for buyer questions the initial evidence missed, then check those questions against customer language and commercial relevance before adding them. Any additions stay separate from the original comparison. Otherwise, changing the list can masquerade as changing performance.
Fix a commercial gap, not the lowest score
The report will tempt you to fix the lowest score first. Do not. A zero on an informational question with no commercial evidence behind it is not automatically a priority. It is a zero.
My fourth prompt asks for a bounded response:
Recommend one page or factual update supported by these findings. Show the buyer question, evidence, proposed change and approval required before publication.
One page. One buyer question. One defensible factual change. For this paid-search-led method, the evidence should identify the converting term, spend and downstream outcome available. The proposed change must name something concrete: price, an implementation step, an integration or a limit. If Claude cannot name the fact, there is nothing to publish yet.
I keep publishing approval on. Claude proposes through groas MCP, I approve, and every change is logged with its reason and can be undone in one step. The point is not to turn a small edit into a committee meeting. It is to stop an inferred answer from becoming a published product claim.
This is the cost of the alternative: inspect the evidence, approve a small watchlist and make a deliberate edit. It requires judgment. A bundle of generated prompts is easier to buy.
A broad prompt-based tracker can produce misleading coverage numbers and invite broad rewrites. A 20-question watchlist tied to conversions invites one edit I can defend: we were missing on this decision question, here is the fact we added, here is the approval log.
Put four things on the 2027 planning slide
My 2027 slide has four lines and no vanity chart:
- The approved watchlist and its sources. Say the account spends $20k a month and 11 search terms produced 34 SQLs in 90 days. Those 11 terms might become nine buyer questions, each listing the originating term, conversion action and spend. Customer evidence can expand the list from there.
- The commercially relevant gaps, by assistant. Lead with the questions where competitors appear and the brand does not. Keep recommendation gaps separate rather than blending every miss into one score.
- The action taken. Name the page, the fact added, the approver and the change log. “Improved AI visibility” is not a description of work.
- The limits still standing. Include thin SQL volume and questions paid search never saw. An evidence-led list is not a claim to know the whole market.
The line I refuse to put on that slide is revenue attributed to share of voice.
Share of voice counts appearances in prompts I chose. It does not count buyers, clicks or pipeline, and it moves when I change the list. I report it as a before-and-after diagnostic on a watchlist I can defend, collected under stated conditions.
Translate that lift into attributable revenue and a useful diagnostic becomes a fiction finance will remember. The converting searches justify which questions to investigate. They do not turn every later brand mention into a sale.
Skip the ad-account start when the receipts are bad
This method fails without useful conversion evidence. If the account has fewer than a dozen real conversions in the last 90 days, its primary conversion is still a pageview or newsletter signup, or its CRM import never worked, I would not build the watchlist around that search-term report.
Bad conversion tracking gives the watchlist the same bias as the bidding. If Smart Bidding is learning to buy cheap fills, a list built from those fills does not become commercially sound because Claude rewrites it elegantly.
For B2B accounts, the stronger foundation connects MQL, SQL or closed-won outcomes back to the click. Without that connection, I fix tracking first or set the paid-search-led approach aside for a quarter. I do not promote weak signals into proof to keep the project moving.
The principle still holds: evidence before prompt volume. I start instead with ten documented customer questions from sales calls, support tickets and GSC question queries with impressions, then add relevant Reddit objections to build a sourced watchlist of around 15 questions. I preserve the original language where available and label any paraphrases.
Those questions can earn a place without ad spend behind them. They simply cannot be presented as conversion-backed. On the planning slide, I name the source each question actually has, not the source I wish it had.
Paid media is a fast source of evidence. Paying for media is not evidence by itself.
A bigger report can send the work in the wrong direction
Vendors sell breadth because breadth is easy to count. A 1,000-prompt report gives the boss a chart with green arrows and gives the team cover in the meeting. It also trains everyone to chase the average.
The mechanism is straightforward: fill the watchlist with informational questions, score them alongside buying questions, then prioritize whatever lifts the blended number. The report can improve while the alternatives and pricing answers still send buyers to competitors. The list looks complete. The work goes where the buyers are not.
Keep the first pass small enough to finish:
- Pull 90 days of converting search terms and inspect the conversion quality.
- Group them by buyer decision and write 9 to 15 questions in customer language.
- Lock the competitor set and check mentions and citations separately by assistant.
- Use customer evidence and competitor citations to propose additions, without rewriting the original baseline.
- Ship one factual fix, approve it and log it.
That is a watchlist I can defend. Not because 15 is a magic number, but because each question has a reason to be there and each proposed edit has evidence behind it.
Skip the bundle that leads with prompt volume and calls it strategy. Walk into 2027 planning with the wrong list and the consequence is not merely a misleading slide. It is a team spending its time improving answers that never mattered while competitors take the shortlist questions nobody thought to track.
Coverage reports comfort. Receipts decide.

