A wave of GEO tools and consultants are telling brands the same story: you need to track thousands of prompts across every LLM, every day, forever — or you’re “invisible” in AI search. The pricing follows the pitch. Dashboards with prompt counters ticking upward. Alerts for every fluctuation. A subscription that scales with prompt volume, so more prompts means more revenue for the vendor, whether or not it means more insight for you.
This isn’t a neutral recommendation. It’s a business model borrowed wholesale from rank-tracking SEO tools, applied to a technology that doesn’t behave the same way. Before you sign a contract sized around “thousands of prompts,” it’s worth understanding why that number is mostly arbitrary and what a defensible, right-sized measurement practice actually looks like.
Why Prompt Tracking Isn’t Rank Tracking
Rank tracking works because the underlying object is stable. A keyword like “best running shoes” has a search volume, a SERP, and a position. Track the same keyword daily and the position moves for real, attributable reasons — an algorithm update, a competitor’s new backlink, a content refresh. The keyword itself doesn’t change meaning between checks.
A prompt doesn’t work like that, for five structural reasons.
1. Personalization. The same question can return different answers depending on who’s asking, whether they’re logged in, and what persona the model infers. A “best CRM for a startup” prompt run as an anonymous session isn’t necessarily the same measurement as the same prompt run against a developer’s account history. Rank trackers don’t have this problem — a SERP position is the same for everyone in a given location. A prompt’s answer set isn’t.
To be fair to traditional SEO, rank tracking has dealt with personalization before. A client’s sales manager once searched for their biggest competitor daily — no bookmark, just habit and Google’s logged-in results quietly started ranking that competitor higher for him personally. Run the same search in an incognito window and the “true” organic position reappeared. The SEO industry solved this decades ago with a convention: incognito, logged-out, geo-controlled — that’s the baseline everyone measures against, and a personalized result is treated as the anomaly, not the ground truth.
GEO doesn’t have that convention yet, and it can’t simply borrow it. Logging out of ChatGPT strips account history the same way incognito strips cookies — but it doesn’t strip persona, because the model can infer who’s asking from the prompt’s own wording, in a single turn, with no login at all. “I’m a developer evaluating CRMs” and “I’m a CMO evaluating CRMs” will shift the answer even in two back-to-back anonymous sessions. Part of GEO’s personalization problem is session-based, like the manager’s search history, and that part is controllable the same way rank trackers control for it. But part of it is now baked into the prompt itself, and no logged-out convention fixes that. Until the industry agrees on what “true position” even means for a prompt — and it may need more than one convention, not the single incognito fix that worked for search — brand-vs-brand comparisons across tools and reports won’t be measuring the same thing.
2. Context and session history. LLMs carry conversational state. A prompt run cold produces a different answer than the same prompt run after three prior turns about budget constraints or prior brand mentions. There’s no such thing as a truly “clean” prompt run unless you control for session state every single time — and most people running these audits aren’t.
3. Intent variability. “What’s the best noise-canceling headphone under $300?” and “Which headphones should I buy for under $300 that block noise well?” are the same intent, phrased two ways. A keyword tool treats these as two distinct queries with two distinct volumes. An LLM treats them as functionally the same request — but may still answer them differently, because small wording changes shift which training associations or retrieved sources get activated. This means prompt count isn’t a good proxy for coverage. Ten well-designed prompts covering ten distinct scenarios tell you more than a hundred paraphrases of the same three scenarios.
There’s actual data behind that last claim, and it points to a useful parallel with search history. Before Google’s Hummingbird update in 2013, keyword tools would often group synonyms and plural variants, “shoe” and “shoes,” “cheap” and affordable” into one bucket for counting purposes, because the search engine itself was starting to treat them as the same underlying question. A 2026 study from Peec AI, analyzing nearly 38,000 AI responses across five engines, found something similar happening naturally with prompts: when real people phrase the same underlying question in their own words, those different phrasings turn out to be far more alike than you’d guess — roughly 9 out of 10 phrasing variations landed close enough in meaning to count as effectively the same question, with only a small slice drifting far enough to become a genuinely different question in the model’s eyes. And which brands showed up in the answer barely changed as long as the phrasing stayed in that “same question” zone — the only real drop-off happened when wording drifted so far it became a different question outright, at which point brand visibility fell by roughly half.
That’s the modern equivalent of Hummingbird’s synonym clustering, and it means a single well-chosen prompt can do double or triple duty as a stand-in for several close phrasings, rather than needing a separate test for each one. It also opens a technique worth exploring directly: testing deliberately broader, more general prompts that are written to capture the territory several narrower variations would occupy, and treating a strong result there as a reasonable proxy for how those narrower phrasings would likely perform too — rather than testing every specific variation individually. That’s a meaningfully different discipline than the “peanut-buttering” approach earlier in this piece, though: it only works if the broader prompt is deliberately built to sit inside the same tight similarity cluster as the phrasings it’s meant to represent, not just a vaguer, unrelated version of the topic.
The one place this doesn’t hold: the same research found that unbranded, comparison-style questions in the middle of the buying journey are far more sensitive to exact wording than top-of-funnel or branded questions — small changes in phrasing there can genuinely change which brands the AI surfaces. That’s the zone where paraphrase count still matters and a single proxy prompt isn’t a safe substitute for testing the real variations.
4. Output non-determinism. Run the identical prompt five times and you can get five different answer sets, purely from sampling variance. A rank tracker checking the same keyword twice in an hour will get the same SERP (barring a live algorithm event). A prompt checked twice in an hour might legitimately return different competitor sets. This is why a single prompt run is close to meaningless — but it also means adding more distinct prompts doesn’t fix the noise problem. You fix noise with repetition on the same prompt, not with volume of different prompts.
5. Judgment-based scoring. A SERP position is a fact: you’re #3 or you’re not. An AI answer requires human judgment to score — was your brand the primary recommendation, a footnote, or framed negatively? That scoring doesn’t automate cleanly, which means every additional prompt in your tracking set adds real human review cost, not just a line in a database. This is the piece most “track thousands of prompts” tools quietly skip: they report presence/absence, not the framing and weighting that actually matter to a buyer’s decision.
Put together, this means volume is a false signal of rigor. A vendor tracking 3,000 prompts a week without addressing session state, without re-running for stability, and without human scoring for framing has produced a large amount of noise, not a large amount of insight.
Where Do You Even Get the Prompts?
There’s a question that comes up in nearly every GEO call and workshop, almost always early: “Where can I get my prompts?” It’s usually asked the way people ask for a report — as if there’s a button somewhere that exports the list.
That expectation exists because paid search and Google Search Console trained a generation of marketers to think keyword discovery is a data problem. Click a few buttons, get a list of queries, sort by volume, done. There is no equivalent for prompts. No engine hands you a report of what people are asking about your category. And the GEO tools rushing to fill that gap are, in large part, selling the same promise the early keyword tools sold: point us at your site and your competitors, and we’ll puke out a list of “related prompts” for you.
This is deja vu from the early days of search — except with one difference that makes it worse, not better. In the early keyword era, at least nobody expected a list to already exist. You built it with judgment: customer interviews, sales call transcripts, support tickets, category awareness levels. Building the list was the strategy work. GEO arrived into a market that had already been spoiled by twenty years of GSC and paid search dashboards, so the instinct now is to skip straight to the tool and treat prompt discovery as a solved, automatable step — when in most categories it’s closer to where keyword research was before anyone had built the dashboards at all.
A few years ago, a multinational client wanted to build a keyword opportunity model from scratch. We started with a six-slide questionnaire designed to force the room to think through customer needs, category awareness, and competitive gaps before touching any tool. Two slides in, someone raised a hand: “I found software that can do this for us — it’ll pull all the words related to us and our competitors. Can we just use that instead?”
The room wanted the shortcut because the questionnaire was uncomfortable. It forced them to admit that most of their existing content didn’t match what people were actually searching for. The head of digital told them to push through anyway. By the end of that one session, manually, we had over 100 content ideas and just as many identified gaps — and everyone in the room understood, for the first time, how their buyers searched differently depending on their awareness of the category, the problem, and the brand. The content and brand teams left less than thrilled, because the exercise surfaced how much of their work missed the mark. The product and business-insight teams loved it, because it was the first time search behavior had been tied directly to what customers actually needed.
The same dynamic is playing out with prompts. A tool that auto-generates a prompt list will give you volume, and it will feel like progress. What it won’t give you is the thing that made that workshop valuable — the forced, uncomfortable process of mapping prompts to real awareness stages, real objections, and real gaps in what you currently say about yourself. That process doesn’t shortcut well, then or now. If there’s a first step to a GEO prompt program, it’s not sourcing a tool — it’s running your own version of that six-slide exercise before you let software fill in the rest.
The Fix: Mini Indexes, Not a Single Visibility Score
There’s a related critique worth folding in, made recently by Seth Besmertnik, CEO of Conductor, in a LinkedIn post, and it points at the same flaw from a different angle: the “visibility score” most AEO platforms lead with — a single unified number for your whole AI search performance — is close to meaningless, not because measurement itself is impossible, but because a single score papers over the exact variables we’ve been describing. His argument, in short: AI search visibility scores in most popular AEO tools today aren’t reliable, but AI search is very measurable once you get focused enough. The inconsistency people notice between tools isn’t proof the category can’t be measured — it’s proof that most companies aren’t measuring it narrowly enough, and are then reporting an unreliable number up to leadership as if it were a KPI.
His point sharpens something we only gestured at with personalization: a single keyword search in the old world had one intent, one SERP, and everyone roughly saw the same results. A prompt doesn’t — the same topic can be asked by a dozen different personas, at different stages of awareness, for branded and non-branded reasons, and each of those combinations can produce a meaningfully different answer set. A blended “visibility score” averages across all of that variation and hands you a number that isn’t tied to any real decision anyone is making.
His fix is to build mini custom prompt indexes — narrow, deliberately constrained clusters, each defined by:
- 1–3 specific personas, written out as actual paragraph definitions rather than a role label, since a prompt set that isn’t persona-driven is of limited use — everyone asking has some implicit persona whether you define it or not. Include proxy searchers where relevant, not just the end user or economic buyer: a caregiver researching on behalf of a patient, or an IT admin evaluating on behalf of an exec, phrases things differently from the person who’ll actually use the product, and a persona set that only covers “the buyer” can miss where a real chunk of the prompt volume is coming from
- A single customer journey stage, but defined more precisely than “awareness, comparison, decision.” The sharper cut is whether the searcher is problem-driven (they know something’s wrong but don’t yet know a category of solution exists — “why won’t this invoice get approved on time”) or solution-driven (they know the category and are now comparing or validating within it — “best AP automation software for mid-sized companies”). Those two searchers want fundamentally different things back — the first wants education and validation, the second wants options and proof — and blending them in one index will muddy the read the same way blending personas does.
- A specific topic or product area, narrow enough that the prompts inside it are actually comparable to each other
Besmertnik’s own example, from Conductor’s AEO practice: F500 CMO, Enterprise VP of Digital Marketing, and Head of Content & AEO as personas; “comparison” as the journey stage; and an “AEO shortlist” topic — essentially the prompts a buyer would use to build an RFP shortlist. Run consistently, that narrow index produces something a blended score can’t: a stable read on which brands get recommended, how they’re framed, and which citations are driving it, with enough week-to-week consistency to actually trust the trend line.
This also surfaces a distinction the tiered framework above doesn’t call out explicitly: branded and non-branded prompts aren’t the same exercise. A prompt that names your brand (“Is [Brand] a leader in AEO?”) is testing entity understanding. A prompt that never mentions you (“best AEO shortlist tools for enterprise”) is testing whether you get pulled into consideration at all. Blending both into one index, or one score, muddies which problem you’re actually diagnosing.
The practical implication for the baseline below: don’t spread your prompts evenly across every persona and topic you can think of — what’s sometimes called “peanut-buttering” prompt generation. Most teams only have the resources to actually act on a handful of focus areas at a time, so build one or two mini indexes deep and specific first, prove out the process, and expand as your capacity to act on the findings grows. A wide, shallow set of prompts covering everyone and nothing in particular is the AEO equivalent of a visibility score — it looks comprehensive and tells you almost nothing you can act on.
So How Many Prompts Do You Actually Need?
Fewer than you’ve been told, and better constructed than a scraped keyword list. Deeper dive in this article on How Many Prompts Do you Need?
Start with coverage, not count. The goal is representative coverage of the scenarios, brand questions, and product battles that matter — not exhaustive coverage of every phrasing. A reasonable starting set:
- 10–15 Tier 1 (Intent/Scenario) prompts — the real jobs-to-be-done your buyers bring to AI, built with a consistent structure (role, objective, constraints, output format) so results are comparable over time.
- 8–12 Tier 2 (Brand Understanding) prompts — identity, expertise, competitors, authority. This tier is small because there are only so many ways to ask “what is this brand and is it credible,” and it doesn’t need to scale with your product catalog.
- 10–15 Tier 3 (Product) prompts — comparisons, use-case fit, and price/value questions for your highest-revenue or highest-competition SKUs, not your entire catalog.
That’s roughly 30–40 prompts total to establish a baseline — not thousands. The rest of the rigor comes from how you run them, not how many you run:
- Repeat each prompt 3–5 times across separate sessions to get a stability score, rather than adding new prompts to pad the set.
- Run the same set across the 3–4 engines your buyers actually use, since sourcing behavior differs meaningfully between them.
- Re-test on a fixed cadence (quarterly is reasonable for most categories; monthly only if you’re in a fast-moving, heavily contested category) rather than continuously — continuous tracking of a non-deterministic system mostly generates alert fatigue.
- Score with a human rater for framing and positioning, not just presence/absence, because a neutral or negative frame is a materially different problem than simple invisibility.
Scaling beyond this baseline is a legitimate next step — but it should be driven by what the baseline reveals (a specific product line losing ground, a specific persona getting worse answers) rather than by a vendor’s default package size. Expanding prompt volume without expanding review capacity just produces a bigger pile of unscored data.
The Honest Trade-off
There’s a real tension worth naming plainly: more prompts, more engines, and more repetitions genuinely do surface things a small set will miss — a competitor quietly winning one persona, a product category where you’re invisible only on one engine. Vendors selling volume aren’t wrong that scale reveals more. They’re wrong to imply that scale is the first problem to solve, or that it substitutes for the harder, less automatable work of scoring how you’re actually framed.
A 30–40 prompt baseline, run with discipline around repetition and human scoring, will tell a mid-sized brand almost everything it needs to know to decide where to invest next — in structured data, in analyst coverage, in content that ties your brand to the attributes AI engines associate with your category. That decision is worth making before committing budget sized around a prompt count someone else picked for you.
