Overview

AI assistants are shaping opinions across every marketing and growth channel, and they are increasingly making purchasing decisions on behalf of your buyers.

This guide outlines how we measure AI attribution at Unusual. It is written for marketing, growth, and revenue teams who need to measure AI's effect on their pipeline.

Four questions to answer

After running this work for dozens of B2B companies, we've landed on the same four questions every time. They map cleanly to the parts of the funnel a revenue team already cares about, and they replace the vague "how do we measure AI" with something a CMO can answer each quarter.

1. What is AI's perception of us? (brand perception)

2. Is AI bringing us leads? (lead origination)

3. Is AI making our leads more likely to buy, faster? (pipeline conversion)

4. Is AI acting as a customer itself? (agents as buyers)

The rest of this guide walks through each one: the signals worth running, the rough order to run them in, and where the framework breaks down. A few practical points before we start.

AI is two things at once. It's a channel that sends some buyers directly to you, and it's an influence on every other channel you run. The two roles require different measurement. The channel side is the small, visible piece; the influence side is the larger, harder piece. A serious attribution program measures both.

No single signal answers any of these questions. Self-reports are biased, server logs show attention without opinion, branded search is noisy, cohort tests are directional. Triangulation is the approach. Collect several signals of varying strength on each question and read them together. When they agree, the answer is clear. When they disagree, the disagreement itself is diagnostic, and it points to what you should instrument next.

Question 1: What is AI's perception of us?

Before any of the channel questions are worth asking, you need to know whether AI has a view of you at all, and whether that view is favorable. This is the perception layer, and it's the foundation the other three questions sit on top of.

The shape of the measurement is closer to qualitative customer research than to analytics. For each persona you sell to, and each buying context that persona enters AI conversations with (a procurement evaluation, a midnight Slack thread about an integration headache, a board prep), you sample AI conversations across the major models and aggregate the output along three axes.

Routing. How reliably do AI models route or associate buyer problems with your solution category? When a buyer describes their problem in plain language, does AI hand it to the category you compete in, or to a different one? If the problem routes somewhere else, AI's view of you inside your category never reaches the buyer.

Visibility. How reliably do AI models find your content when they look for solutions within your category? Essentially the SEO question for AI: when AI is researching unbranded queries inside your category, does it actually retrieve your pages, your reviews, and the rest of the surfaces you've published on?

Alignment. An LLM's overall opinion of you. When AI considers the options inside your category, does it believe you're the best solution compared to competitors, and does it describe you that way? This shows up in the direct read (how AI characterizes your brand on its own) and the comparative read (how AI weighs you against named competitors). Alignment is the dimension where positioning, proof, and content investments do the most work.

Underneath those three, you're tracking two more things continuously. Which framings AI uses about you (the reflected language), and which pages, reviews, and external sources AI relies on to construct its answer (the citations and sources). Both sit inside the per-cell measurement as components that explain why the three axes land where they do.

The most useful thing this read produces is the diff against the prior period after a content investment shipped, a competitor moved, or a model retrained. Absolute scores on any one cell matter less than movement. When several cells move together after a coordinated piece of work, you have a clean line between content and observable change in market perception.

Supporting signals that triangulate the platform read:

• Pitch-versus-perception cross-reference. Catalog every product claim your sales team asserts on calls. Compare each claim against the latest AI snapshot: repeated, silent, or contradicted. The repeated bucket validates the perception read. The contradicted bucket is your direct content priority list (sales is saying X, AI is saying not-X, the action is the proof page that closes the gap).

• Buyer language in sales calls. When buyers raise unprompted the same caveats AI uses, AI is reaching the market and the framing is surviving into the conversation. When AI framings never appear in buyer language, either AI's influence isn't reaching that persona, or the sampling missed the contexts that matter.

• Post-purchase customer survey. Ask recently converted customers what they asked AI during evaluation, what AI told them about you, and what AI told them about competitors. A small sample is enough to validate or contradict the platform read.

Doing this work well at meaningful scale across personas, models, and buying contexts is the part most teams underestimate. It's why we built Unusual. You can start manually with a handful of representative conversations per persona and a structured tagging pass; the discipline matters more than the volume.

Question 2: Is AI bringing us leads?

The lead-origination question splits cleanly into two halves. The first is direct: how many buyers click through from an AI surface to your site. The second is indirect: how many buyers arrive through every other channel only because AI told them to. The first is small and easy to measure. The second is large and harder.

The direct side

Most teams underbuild this side, which is a shame because it's the cheapest data to capture.

• UTM source from AI surfaces. Configure your analytics to capture `utm_source` values for chatgpt, claude, perplexity, gemini, and copilot. Count and conversion rate of buyers arriving via AI citations or referrals. A hard floor on AI-direct traffic.

• Referrer tracking from AI domains. For visits that don't carry a UTM, capture referrer domain (chatgpt.com, claude.ai, perplexity.ai, gemini.google.com, copilot.microsoft.com). Catches a meaningful share of AI clicks that lose the UTM in transit.

• Bot user-agent filtering. Real-time AI consultation on behalf of an active buyer registers as a visit from `ChatGPT-User`, `Claude-User`, or similar. Volume tells you the AI attention shape over time. Concentration (which pages bots visit) tells you what AI is asking about you on behalf of buyers: if visits land on your integrations page, buyers are asking AI about integrations; if they land on a competitor comparison post you wrote, buyers are researching that competitor and you're the nearest authoritative source. If your site is behind Cloudflare, our [Agent Analytics](https://analytics.unusual.ai) tool does this out of the box. It's free, open source, and tracks fourteen AI crawlers across the major labs.

• Form-fill self-report. "How did you hear about us?" with AI as a dropdown option, or a dedicated AI question on the lead form. Biased toward buyers who consciously noticed AI use and are comfortable saying so. Some enterprise customers won't add another field to their lead form; use this where you can.

The indirect side

This is where the volume lives, and where attribution gets harder. The buyer who arrives through paid search after a thirty-minute Claude conversation about you. The buyer who type-ins your domain after AI mentioned you favorably in a chat they never clicked through from. The buyer who searches your brand on Google because a friend recommended you, and the friend got the recommendation from ChatGPT. None of this shows up in your channel attribution, and all of it is moving who enters your pipeline.

The signals here are correlational by nature, so the work is reading them together rather than relying on any one.

• Direct and type-in traffic correlation with content drops. Overlay direct-traffic spike windows against publication timestamps of new content and platform-detected perception movements. When AI conversations about you intensify, direct traffic to your domain often follows within days. Confounded by every other thing that drives direct traffic (campaigns, events, news), so read with branded search for the cross-check.

• Branded search lift. Branded search volume from Search Console with intervention timestamps overlaid. Lifts after content drops or perception movements indicate AI conversations the platform doesn't directly sample, which is most of them. Free, fast-moving, already collected.

• Page-level bot visit topic concentration. The same bot data, cut by page topic. Which pages on your site are AI agents actually reading on behalf of buyers, and how does that concentration shift over time? A proxy for what AI is asking on your behalf during real buyer conversations.

• Long-tail unbranded query trends. The emergence of conversational, AI-style queries (long phrases, specific scenarios, comparison patterns) in your organic search data is a leading indicator that the buying market is shifting to AI-mediated discovery in your category. Useful as supporting context rather than as a load-bearing measurement.

Question 3: Is AI making our leads more likely to buy, faster?

This is the question revenue teams care about most. By the time AI is influencing your pipeline meaningfully, the question shifts from "is it happening" to "is it helping or hurting." The cleanest answer comes from tagging the AI-affected cohort and comparing its pipeline performance against the rest of your leads on conversion rate, cycle time, average contract value, and where it falls out of the funnel.

This question has its own deep-dive companion guide, because the answer affects investment decisions across content, sales enablement, and product positioning. The short version below; the long version walks through cohort construction, the supporting signals that make the cohort number trustworthy, and what to do with the result. See How to Tell If AI is an Advocate or Detractor for Your Brand?

The core moves:

• AI-affected cohort funnel comparison. Compare an AI-influenced cohort (defined via transcripts, UTM, or any combination of upstream attribution methods) against the baseline cohort on conversion rate by stage, win rate, average deal size, cycle time, and loss-to-competitor pattern. The headline measurement.

• Buyer-belief extraction from transcripts. Tag every claim, objection, and belief the buyer brings into a sales call. Cross-reference each against the platform-tracked AI framing on that dimension. When a buyer's belief matches AI's framing on a dimension you haven't published on, AI is the proximate origin.

• Close-reason cross-reference (won and lost). Take what AI says about you and line it up against the reasons your buyers cite for winning and losing deals. When buyers cite strengths on dimensions AI also describes favorably, AI is reinforcing what's working. When buyers cite weaknesses on dimensions AI warns them about, AI is actively costing you deals. The latter list is your content priority.

• Lost-deal interviews. A rolling cadence of five to fifteen interviews per quarter with closed-lost contacts. The cleanest qualitative attribution available, since buyers have no incentive to hide their reasoning after the deal is closed-lost.

Question 4: Is AI acting as a customer itself?

The last question is about agents themselves becoming buyers. Most teams treat this as future tense, but agents are already booking demos, requesting pricing, and starting trials. The volume is small and growing quickly.

What's worth measuring today is presence and intent. Are agents arriving at your site on someone's behalf, and what are they asking?

• Bot crawl logs cut for active buyer accounts. Filter the bot visit data from Q2 for IP ranges corresponding to active prospect or customer companies. When `ChatGPT-User` visits your pricing page from a known prospect's IP range during their open sales cycle, you have direct evidence that AI was consulted by someone in that account mid-evaluation. Per-deal evidence with timing alignment, useful for narrative attribution on specific deals.

• AgentDesk de-anonymization. Our AgentDesk surface acts as an AI-facing chat that other AI agents query about you when researching you. When an agent visits, the platform captures who the visiting party is (where attribution data is available) and what they asked. Works as a lead-generation surface and as a direct read on what AI is asking on a known account's behalf.

• Agent conversion rate, as it stabilizes. Once agents can book a demo or start a trial on their own, agent conversion rate becomes a pipeline metric you track directly. Today this is a leading indicator. Within a year or two it will be a line item on the same pipeline dashboard you already report on.

This is the question that's going to grow most in importance over the next eighteen months, and the work to instrument it is mostly the work you're already doing for Q2. The Q4 cut is recutting the same data with a different lens.

Signals to ignore

Three metrics show up consistently across AEO and GEO dashboards: share of voice, composite brand scores, and rate metrics on fixed prompt sets. They sound compelling but have structural issues that keep them from answering any of the four questions.

Share of voice. The fraction of AI answers that mention your brand across a hand-picked set of prompts. The problem is the prompts. They're picked by you or your vendor, not by your buyers. Tiny word changes ("best" to "top") swing the number materially. Real buyers ask questions in their own phrasings, in long conversations, with context about themselves the model has and you cannot see. Optimizing against a fixed prompt set improves your score on that prompt set and generally fails to generalize to the conversations that matter. A longer treatment is in How Stable is Share of Voice?

Composite brand scores. A handful of AEO tools roll several underlying metrics into a single composite score, framed as a one-number read on your AI presence. The problem is opacity. If you can't explain what the number measures in terms of buyer behavior or AI belief, you can't optimize against it honestly. You can only move it in whatever direction the tool's hidden weights happen to reward. Any composite score that arrives without a transparent explanation of its inputs and weights is telling you more about the tool than about your brand.

Rate metrics on fixed prompt sets. Citation rate, inclusion rate, recommendation rate. The same prompt-sensitivity problem as share of voice. When the vendor updates its prompt set (and they do, silently), your historical numbers become incomparable. When the underlying models update (and they do constantly), your numbers move for reasons that have nothing to do with your brand. These metrics work as narrow activity indicators within the frame the vendor defined. They have no place in a pipeline-attribution framework.

Running the cycle

The four questions are designed to be answered together, on a quarterly cadence. The shape of the work is the same each time:

1. Pull the current read on each question, using the two or three signals you've decided to run for each one.

2. Cross-reference what AI believes about you against what your buyers are citing in won and lost deals, objections, sales transcripts, and lost-deal interviews.

3. Synthesize. For each persona and buying context, is AI an advocate or a detractor right now, on which dimensions, by how much, with what confidence band?

4. Decide on two or three perceptions worth moving this cycle, and the positioning, content, or proof work to move them.

5. Re-measure next cycle against the same baselines.

The output of every cycle is a named list of perceptions worth moving and a corresponding set of content investments to make. That's the loop. Same shape brand marketers have always used to shift an audience's view of a company, applied to an audience that happens to be everywhere your buyers are and answering their questions in real time.

What comes next

The four questions are four lenses on the same underlying question every revenue leader is asking right now: how much is AI's perception of our brand actually moving our pipeline, for which buyers, and where?

The tools to answer each one exist today. The discipline that turns the signals into decisions is the same discipline brand marketers have always used, applied to an audience you can now finally measure.

We built Unusual to run the perception read on AI brand perception (Question 1) and to triangulate the other three against it across personas, models, and buying contexts. If you want to see what your read looks like, that's the place to start.