Stop Tracking Keywords. Start Tracking Prompts.

Open any SEO dashboard and you will find the same artifact: a flat list of keywords, each tagged with its monthly volume, a rank, and one little trend arrow. That list was a reasonable model of the world for about twenty years. A person typed two or three words into a box, ten blue links came back, and you competed for position on those exact strings. The keyword was the atom of demand, and counting keywords counted demand.

That world is draining away. When someone asks an assistant a question, they do not type "crm for startups." They type "we're a 12-person team, mostly technical, moving off spreadsheets, we hate Salesforce, what CRM won't we regret in a year." The keyword list has no row for that. It cannot. And the assistant does something with that sentence that no ten-blue-links engine ever did, which is where the whole model of keyword tracking quietly falls apart.

One prompt is not one query

Feed a conversational question into a retrieval-augmented system and the first thing it does is take your sentence apart. The model rewrites the prompt into several sub-queries, a step usually called fan-out. Our example about the 12-person team might fan out into "best CRM for small technical teams," "CRM alternatives to Salesforce," "CRM with good spreadsheet import," and "CRM pricing for startups under 20 seats." Each sub-query hits a search backend separately. ChatGPT uses provider-managed web retrieval, Perplexity operates its own crawler and a provider-managed retrieval stack, and some systems query a vector store instead. Each retrieval returns its own candidate set. A cross-encoder reranker then trims the pooled results down to a handful of passages, those chunks get assembled into the context window, and only then does the model write an answer and staple citations to it.

Count the queries in that flow. One prompt from a human became four machine queries, four candidate sets, and one synthesized answer drawing from all of them. Your keyword tracker, watching for the string "crm for startups," saw none of it. The words the customer used never appeared in any of the four sub-queries the machine actually ran, and the words the machine ran never appeared in your keyword list.

This is the first crack. A keyword assumes a one-to-one map from what a person types to what an engine searches. Fan-out breaks that map into pieces and scatters them. You can rank first for every keyword in your spreadsheet and still be absent from the four sub-queries that decided this answer, because those sub-queries were generated on the fly from a sentence no spreadsheet contains.

The long tail stopped being a tail

Keyword thinking has always had a category for weird, specific, low-volume searches. We called it the long tail, and the word "tail" did real work: it implied a small, optional appendage hanging off the fat head of high-volume terms. You chased the head first and picked at the tail if you had spare time.

Conversational input inverts the shape. When the interface is a chat box instead of a search bar, people stop compressing. Nobody edits a spoken-style question down to three keywords, because the assistant rewards detail rather than punishing it. So the input arrives long, clause-heavy, and loaded with context: the team size, the incumbent tool they resent, the specific fear, the budget ceiling. Almost every one of those sentences is, by keyword-volume standards, a zero-volume string. It has never been typed before in that exact form and never will be again.

A flat keyword list handles zero-volume strings by ignoring them, since volume is the only thing it knows how to rank by. But the "tail" is now the whole animal. The high-volume head terms still exist, and they still matter for classic search, yet a growing share of high-intent, ready-to-buy questions live entirely in phrasings your volume tool scores as noise. You cannot prioritize by volume when the valuable queries all have a volume of one.

Intent clusters, not strings

Here is the move that gets you out of the trap. Stop treating each phrasing as a distinct thing to track and start treating phrasings as surface variations on a smaller set of underlying intents.

Our four sub-queries from earlier were not four topics. They were one buyer, in one situation, expressing one intent from four angles: replace an incumbent, fit a small technical team, import existing data, stay under a budget. Two hundred different humans will phrase that same buying situation two hundred different ways, and the assistant will fan each phrasing out differently again. The strings are effectively infinite. The intents behind them are not. In most categories the real intent space is countable: a few dozen distinct jobs a customer is trying to do, each of which shows up in comparison-shaped, problem-shaped, and category-shaped variants.

That countable set of intents, mapped across the phrasings and sub-queries that express them, is what we call the prompt space. It is the object that replaces the keyword list. A keyword list is a column of strings ranked by volume. A prompt space is a structured map of the questions your customers actually bring to an assistant, grouped by the intent underneath and annotated with how each one fans out.

Prompt-Space Coverage

Once you have the map, the metric follows. Prompt-Space Coverage is your visibility measured across the real space of prompts customers ask, spanning intents and phrasings, rather than a rank against a handful of chosen strings.

The distinction that matters most is that coverage is a map, not a number. A single "AI visibility score" collapses everything into one figure and hides the only thing you can act on: where you show up and where you vanish. Real coverage is read as a grid. Down one axis, the intent clusters. Across the other, the phrasing variants and the engines. Each cell answers a narrow question: for this intent, expressed this way, on this engine, do you appear in the answer, and upstream of that, do you appear in the candidate sets the reranker chose from?

That upstream half is why coverage has to sit on retrieval, not just on the final wording of answers. Whether you got cited is the visible outcome. Whether your content entered the candidate set for a given sub-query is the leading indicator, the thing shareof.ai calls Share of Retrieval. A cell can be dark for two very different reasons: the model retrieved you and declined to cite you, or the model never retrieved you at all. Those failures have opposite fixes, and a coverage map that only watches final answers cannot tell them apart. Reading coverage at the retrieval layer shows you the Retrieval Gap directly, the distance between the pages that could have answered a sub-query and the smaller set the system actually pulled.

Building the prompt inventory

None of this works without a real inventory of prompts, and you will not get one from a keyword tool, because keyword tools were built to report volume on strings, not to reconstruct the situations behind questions. You build the inventory from sources where humans describe their actual problems in their actual words.

Mine your sales calls. Recorded discovery and demo calls are the richest source you own. The questions prospects ask a salesperson are almost identical in shape to the questions they now ask an assistant: long, situational, comparison-heavy. Pull the transcripts, extract every question and objection, and you have the raw material for your comparison and problem-framed clusters.

Mine support tickets. Support inboxes and community threads capture the post-purchase and evaluation-stage questions, phrased with all the specificity of someone who has a concrete problem. These map cleanly to the "does it do X" and "how does it handle Y" intents that assistants field constantly.

Read real assistant logs. Where you can capture them, actual prompts to assistants are the ground truth. If you run any kind of chat or copilot surface, those logs already show you conversational phrasing and follow-up structure. Public prompt-sharing, community posts, and your own team testing questions in ChatGPT and Perplexity fill in more.

Generate the framed variants deliberately. For each intent you extract, write out its predictable shapes. A comparison variant ("X versus Y for a small team"), a problem-framed variant ("we keep hitting X, what fixes it"), a category variant ("best tool for X"), and a switching variant ("moving off X, what should we consider"). This is where you make the long tail tractable, by enumerating the phrasing families rather than trying to list every possible sentence.

Cluster what you gather by intent, not by keyword overlap. Two questions that share zero words can belong to the same cluster, and two questions that share a word can belong to different ones. The output is a few dozen intent clusters, each with a set of representative phrasings and framed variants under it. That is your prompt space, and it is a living document, because new intents surface as your category and your product move.

Measuring coverage against the inventory

With the inventory built, measurement is mechanical. For each representative prompt, run it through the engines that matter to you, the full set being ChatGPT, Claude, Gemini, and Perplexity. Record two things per run. First, did you appear in the answer, and how (named, cited, linked, or merely alluded to). Second, wherever you can observe it, did your content show up in the retrieved candidate set before the model wrote anything.

Because fan-out is real, do not score only the top-line prompt. Where a system exposes its sub-queries or its sources, log those too, so you can see which specific sub-query dropped you. A prompt can cite you through one of its four sub-queries and ignore you on the other three, and the fix lives in the three that failed.

Roll the results up by intent cluster, and the map appears. Some clusters will be strong across every phrasing and every engine. Some will be strong on ChatGPT and dark on Perplexity, which usually points at a crawl or index difference rather than a content one. Some clusters will be dark everywhere, and those are your real gaps, the intents where customers are asking and you are simply not in the room. Prioritize by the value of the intent, not by any volume figure, because volume as you knew it does not exist here.

A worked read might look like this. Your category-shaped prompts ("best X for Y") cover well, because that content already exists and reads like a listicle the reranker loves. Your switching prompts ("moving off competitor Z") are dark, because nobody wrote the third-party comparison pages that answer them, and remember that most brand mentions in AI answers trace to third-party pages rather than your own site. The gap there has nothing to do with ranking or title tags. You close a hole like that by getting the right passages onto the right third-party surfaces, so retrieval can find them.

Two ways of seeing the same demand

Keyword-volume thinking asks: which strings are searched most, and where do I rank on them? It rewards you for owning a short list of high-traffic terms and lets you ignore everything with low volume. It assumes the query the engine runs is the query the human typed.

Prompt-space thinking asks a different question: across the full space of situations my customers bring to an assistant, and across the sub-queries each situation fans out into, where am I present and where am I missing? It rewards breadth of coverage across intents and honesty about the dark cells. It assumes, correctly, that one human prompt becomes many machine queries, and that being visible means being retrievable across all of them.

You can keep your keyword rankings for the search that still runs on ten blue links. That channel is not gone. But for the growing share of demand that arrives as a conversation and gets torn into sub-queries before any content is retrieved, the keyword is the wrong unit, and no amount of volume data will fix a model that miscounts one prompt as one query. Map the prompt space, score your coverage across it, and you are finally measuring the thing that decides whether an assistant recommends you.

A note on measurement

Building and scoring a prompt space by hand is real work, and it decays the moment your category shifts. The mechanics are all reproducible on your own: pull the transcripts, cluster by intent, run the prompts, log the retrievals. Tools like shareof.ai↗ automate the running and the roll-up across ChatGPT, Claude, Gemini, and Perplexity, so the map stays current without a person re-running two hundred prompts every week. The framework stands without any tool, though. The one thing worth doing today, before you buy or build anything, is to open your last twenty sales calls and write down the questions. You will not recognize them as keywords, and that is the entire point.

Primary sources and evidence

• Web search in the OpenAI API↗ — current search-tool behavior and citations.
• OpenAI web crawlers↗ — official user-agent controls for search and training.

Continue through the system

• Your "AI Rank" Is a Vanity Metric
• Measure Your Own Share of Retrieval With the Model APIs
• The 15-Minute AI Visibility Audit You Can Run Right Now

Run the practical layer with the free AI visibility scan, explore the AI search library, or see the AI visibility benchmarks.