AI Share of Voice: Reading Your Competitors Through AI Answers

Buyers now run their first supplier evaluation inside AI-generated answers. That answer pool is a new public intelligence channel, if you read it as signal instead of as a marketing metric.

The evaluation moved, and most teams didn't notice

The first pass of industrial buying used to happen in places you could watch: a trade-show walk, a distributor call, a request for quotation. Increasingly it happens inside an AI conversation. An engineer asks for alternative approaches to a joining process. A procurement lead asks for a shortlist of suppliers for a tolerance class. A plant manager asks who else makes a component that just went on allocation.

Your sales team sees none of this. The CRM lights up later, when a buyer who has already formed a shortlist makes contact. Everything that shaped that shortlist happened upstream, in answers you never saw.

The consequence for CI is easy to miss because it sounds like a marketing topic. This is not about whether you should "do GEO" or optimize your own content. It is about whether you read this channel. Your competitors' positioning work shows up in AI answers before it shows up in your pipeline, and the answer pool is public, queryable, and surprisingly honest about how the market describes each player. That is an intelligence source by any definition.

There is also a second-order effect. Because the evaluation happens earlier, the buyer who finally contacts you arrives better informed and closer to a decision. Teams often misread this as improved lead quality. Partly it is. But it also means the competitive frame was set before you entered the conversation, and resetting a frame late costs far more than shaping it early.

What an answer pool actually contains

Sample the same questions across major AI engines for a month and the answers stop looking like text. They start looking like data. Four kinds of signal are readable:

• Citation sources. Who gets cited, and from what type of source: the vendor's own site, trade media, standards bodies, or forums. A competitor whose citations come mostly from its own site is talking to itself. A competitor newly cited from a standards body has done positioning work you should know about.
• Triggered questions. Which questions a competitor appears on, and which it does not. Showing up on "best suppliers for X" but not on "alternatives to Y" tells you exactly where its content footprint reaches and where it stops.
• Descriptive language. The category and adjectives the answer assigns: low-cost, premium, niche, regional. This is the market's description of your competitor, filtered through the model, and it moves before official messaging does.
• Co-occurrence. Who your competitor gets compared with. The comparison set is the de facto category boundary. When a new name starts appearing next to yours, a buyer somewhere has already made that association.

Each signal can be misread. Citations fluctuate with retrieval changes. Descriptions lag reality in both directions. A single answer is an anecdote, and the discipline in the next section is what turns anecdotes into a series.

A monthly protocol you can run

This protocol needs no special tooling beyond a spreadsheet and calendar discipline. Its power comes from never changing the instrument.

1. Fix the question set. Write questions from your key intelligence topics in three families: category awareness ("Who are the main suppliers of X?"), vendor selection ("Which suppliers handle tolerance class Y in region Z?"), and alternative comparison ("What are the alternatives to Competitor A's product line B?"). Freeze the wording. Questions you rewrite every month measure your rewriting, not the market.
2. Sample on a fixed cadence. Same engines, same phrasing, same day and time window each month. Record whether each answer carries citations and of what type. Consistency beats coverage: fifteen frozen questions asked monthly beat a hundred questions asked once.
3. Normalize the citation pool. Classify every cited source into one bucket: competitor-owned content, trade media, standards or associations, forums and community, or your own content. Compute the share of each bucket. The structure of the pool matters more than any single citation.
4. Compare snapshots, not answers. Only compare the same question set across months. Look first for structural change: a new source type entering the pool, a description word migrating, a shift in co-occurrence. Only then look at individual answers.
5. Log confidence. Questions with no citations, and questions whose answers swing between runs, get flagged. They are noise until they repeat, and they never enter a decision document.

Run the protocol for two months before you draw a single conclusion. The first month gives you a snapshot; the second gives you a direction. Everything before that is calibration.

Reading movement without over-reading noise

AI answers are stochastic, and engines change underneath you. A model update or a new retrieval source shifts everyone's citations at once, which looks exactly like a competitor move if you only watch yourself. The first filter is therefore comparative: if every player's numbers moved together, the engine moved. If one player moved against a stable field, the player moved.

The second filter is duration. A single-month change is a hypothesis. The same change in the same direction across two or three consecutive months is a finding. This patience feels slow, but the channel itself is fast: you are still seeing positioning moves months before they reach a quote request.

The working rules are few and strict. Only same-direction movement across consecutive months raises a flag. Single-month spikes get logged, not acted on. Engine-wide shifts get recorded as instrument changes, not market events. And no single sampling run ever enters a decision document; the series decides, not the snapshot.

Teams that run this protocol for a quarter end up with something rare: an early view of how competitors want to be seen, expressed in the market's own vocabulary, updated monthly, at almost no cost. That is what a channel is for.