A citation count cannot be converted into money, and most AI visibility dashboards report little else. The chain that ends in revenue has four links: an answer that includes you, a session carrying a readable referrer, an engaged session that fires a key event, and a pipeline record that survives to closed revenue. Three are ordinary web analytics you already run. The second is broken for a large share of AI traffic, and no vendor has repaired it.
The honest answer to "what did our GEO work return" is a range with a named gap in it. Measure answer presence as a base rate across repeated runs, measure the referred sessions you can still see with GA4's session key event rate, which divides sessions with a key event by total sessions, and label everything between as an estimate. The stakes justify the work: G2 found AI chatbots and review sites are now the two biggest influences on B2B shortlists, 51% of software buyers start research with an AI chatbot more often than with Google, and a third said an AI chatbot led them to choose a vendor they hadn't considered.
What is the chain from an AI answer to revenue?
Four links, each measured with a different instrument, and the ROI question goes wrong wherever somebody blends them. A dashboard that multiplies all four produces a number nobody should have to defend in a budget meeting.
Presence is a probability estimated from repeated runs. Referral is a count from your analytics with a known undercount inside it. Engagement is a rate against key events you defined before the reporting period started. Pipeline is a currency figure from your CRM, joined to sessions by whatever identity you can honestly establish.
The top of the chain matters even while the middle leaks. 6sense found that by the time a shortlist is filled, the average buying group member knows 75% of the vendors on it. An answer that reaches a buyer while that set is forming does its work before any session you record.
Where does the measurement chain actually break?
At the join between the answer and the session. A browser passes a referrer only when the linking surface sends one, and a large share of AI answers reach buyers through surfaces that do not: desktop and mobile apps, links copied into a message, and answers read without any click at all. The visit still arrives. Its origin does not.
Analytics then does the reasonable thing and files it under direct. Google's default channel group rules assign each session by referrer and campaign parameters, so a session carrying neither is indistinguishable from someone typing your domain into the address bar.
Two consequences belong in writing on every report you send. Your AI channel number is a floor, and the true total is larger. The direct bucket has stopped being a clean read on brand demand, which matters because 6sense has been tracking where B2B sites are losing traffic to LLMs. Some of that loss is plausibly returning through an unmarked door and being credited to the wrong channel, an inference of ours that 6sense does not make.
How do you configure analytics so the visible half is countable?
Build a custom channel group before anyone argues about attribution. Group the AI hostnames appearing in your referrer report into one channel, keep it separate from organic search, and leave the default groups intact so historical comparisons work. It is an afternoon of setup and the only deterministic part of GEO measurement.
Concretely: a channel named AI Assistants, with source matching chatgpt.com, perplexity.ai, copilot.microsoft.com, claude.ai, or gemini.google.com. Leave google.com and bing.com where they are so organic search keeps its history, and keep AI-adjacent publishers out, because a link inside somebody's article is a different event from an answer. Review the report monthly, add hostnames as they surface, and record the date each joined, so nobody reads a definition change as growth.
Define key events a real buyer would trigger: a trial start, a demo request, a pricing page reached with depth, a documentation signup. Read the new channel through key event rate; raw session counts make a small, high-intent channel look like failure in its first quarter.
A month of AI sessions is often a few dozen visits, and a key event rate computed on forty of them swings several points when two people change their minds. Publish the denominator beside the rate, and refuse a trend line until it earns one.
Which AI visibility numbers are not revenue?
Impressions, mention counts, citation counts, and share of voice are leading indicators at best, and none of them is currency.
A mention is your name appearing in an answer. A citation is your page used as a source. A recommendation is the engine telling the buyer to choose you. Those rungs convert at different rates, so folding them into one visibility number throws away the only information you had.
Composite scores fail the audit for the same reason. Nobody can tell you afterwards which input moved, the weights are a vendor's private opinion, and a score that drops gives you no instruction. The IAB's guidance on measuring visibility in the AI era keeps the dimensions apart on purpose, and names presence, prominence, portrayal, and persuasion as separate things to observe.
The tactics sold against those numbers deserve the same scrutiny. C-SEO Bench, published at NeurIPS 2025, tested conversational-SEO methods across two tasks and six domains and found most of them largely ineffective and often harmful to citation rank, with gains shrinking as more sites adopted them. Meanwhile more than 4 in 10 AI citations come from sources outside the brand's own domain, so much of the work that moves answers happens off your pages.
How do you measure content ROI when AI answers absorb the click?
You price the input and judge it against presence, because there is no session to value. Pew Research Center found users who saw an AI summary clicked a traditional result in 8% of visits against 15% without one, and clicked a link inside the summary in 1%, in browsing data from 900 US adults in March 2025. That is an association, not proof the summary caused the drop.
The tempting fix is to price citations like media: assign a CPM to AI impressions, multiply by citation count, report a dollar figure. Reject it. No engine publishes a per-brand impression count for its answers, and nobody transacts at that price, so the result is invented currency.
Cost and presence are what go on both sides of the ratio. Total the fully loaded program cost, then divide it by the tracked buyer questions where you hold presence at a measured rate. Both inputs are observed, so nothing is inferred: cost per question held is a real number with zero sessions attached. Report it beside the referral floor.
How do you measure the answer layer without fooling yourself?
Start from base rates. Ask each buyer question in a few phrasings on at least two engines, every week, recording how often each vendor appears. SparkToro and Gumshoe's research found AI engines are highly inconsistent when recommending brands, and Search Engine Land's write-up of the same study reports that AI recommendation lists rarely repeat exactly.
Put the uncertainty on the face of the number. A 2026 variance-components study decomposes how much of an observed presence difference is run-to-run noise, and work on quantifying uncertainty in AI visibility makes the case for confidence intervals over point estimates, with platform medians that vary by engine. A presence number without an interval is a claim you cannot check next month.
Semrush's AI Visibility Index analyzed 126 million AI search prompts, six orders of magnitude more sampling than the panel most teams run. Your fifty prompts describe your category only if you chose them from real buyer language.
How do you report this to a budget holder?
In three tiers, labeled. Observed: presence rates with run counts, AI-channel sessions, key event rate, cost per question held, and any pipeline record carrying a readable source. Inferred: the share of direct traffic your referrer gap can account for, stated as a range with its assumption named. Unknown: influence on buyers who read an answer and never clicked, real and unquantifiable this quarter.
Fill part of the unknown tier with self-reported attribution: add a "how did you first hear about us" field to demo forms and read it monthly. It is imprecise and biased toward recency, and it still beats an inference stack assigning credit by arithmetic. 6sense's 2025 B2B Buyer Experience Report is the reason to bother: most of the buying journey happens before you are told it started.
Nobody can guarantee a citation, a ranking, or a recommendation, because the engines are probabilistic and competitors are publishing against you. What a measurement program can promise is a defensible base rate, a channel report with its undercount named, and a rerun after every change.
How does Trovance measure the part of this chain it can see?
Trovance instruments the top of the chain, the part your analytics cannot reach. You define the buyer questions that decide your category, and it sweeps them across engines every week, preserving each answer run with its full context: who was mentioned, who was cited, who was recommended, and which sources carried the answer. Presence becomes a rate with a run count behind it.
Because every answer snapshot keeps its citations, share of answer can be compared across time. When you publish something and presence changes, the next analysis cycle reruns the same questions and shows whether the change held across runs or dissolved into variance. That comparison is what a budget holder is asking for: did the thing we shipped move the answer, and did the movement survive a second look.
The recommended work is scoped to evidence. The Brain keeps the claims you are entitled to make with the proof behind each one, and recommended actions arrive as a ranked queue of content worth making, backed by answer evidence: the benchmark that would counter a competitor's quoted study, the comparison page a third-party source will never write for you. Drafts are written in your voice from your author voice guidance and research sources, and a person reviews and approves everything before it publishes.
What Trovance will not do is hand you a single AI visibility score, promise a citation or a recommendation, or take credit for revenue it cannot trace. The referrer gap described above sits outside any vendor's control, and we are building toward tighter joins between observed answers and observed sessions. Until that exists, you get two honest measurements with the gap between them named. A single confident number would only hide it.
What should you do this week?
Do the deterministic work first: build the AI channel group and define the key events, because every later argument depends on them. Then write the ten buyer questions your category is decided by, ask each in two or three phrasings on two engines every week, and record appearance rates with the answer count. You now have a floor on referrals and a base rate on presence.
Presence can shift once engines re-fetch your pages when the fix is retrieval or evidence on your own pages, and takes longer when it depends on third-party sources carrying you. Pipeline lags both by your sales cycle, so anyone quoting a payback period for GEO in weeks is quoting a number the measurement cannot produce.
If you want the presence half of that chain measured every week, start a free Trovance trial and see what your buyer questions currently return.
Measure the right thing
Does AI search visibility drive leads or revenue? - the evidence on what visibility converts into.
There is no such thing as an AI visibility score - why composite indexes fail an audit.
How to measure AI search visibility without one score - the separate metrics to report instead.
AI share of voice - what the metric can and cannot tell a budget holder.
Track the chain
How to track AI citations - the mechanics of recording answer runs over time.
Why AI citations drop - reading a decline before you spend against it.
How long until AI cites you - realistic timing between publishing and presence.
A crawl is not an audience - why bot traffic is not a demand signal.
