How I Decide If AI Visibility Is Worth the Spend

I fund AI visibility when a number I can re-pull moves a pipeline number the company already believes. Share of citation, the conversion of AI-referred visits, the stability of a mention across repeated runs, the ratio of earned citations to owned ones, and a line from the visit to a pipeline event are the five I will sit in a budget meeting with. A rank, a click, and a session duration are the wrong instruments for a channel that never sent the click.

The dashboard that lays those five out for an operator is How to Measure AI Visibility ROI: The CMO Dashboard. That page is the deeper method. This essay is how I would use it as a founder: which of the five I require before I add spend, what I will accept as a proxy, and the result that would make me pull the budget back.

Why the old dashboard stays green while the answer moves

Forrester's April 2026 argument names the failure directly. The ROI problem is a measurement problem. Teams evaluate AI-driven outcomes with click-through rate, results-page position, page views, and session duration. None of those record whether an engine cited the brand inside a synthesized answer.

Forrester's AI Value Matrix separates value into nine categories on two axes, financial outcomes (revenue, cost, risk) and value mechanisms (productivity, engagement, strategy). Productivity value shows up fast. Engagement value takes longer. Strategic value is the slowest and the most durable. Measuring all three on one timeline is why AI ROI looks inconsistent from quarter to quarter. I put citation work in the engagement-to-strategic range. The attribution chain is longer than a paid click. That is a property of the chain, not a reason to call the work unmeasurable.

Search itself is not the thing that died. Google Search queries hit an all-time high in the first quarter of 2026. Alphabet reported $60.4 billion in search revenue, up 19% year over year. Forrester's February 2026 Consumer Pulse found that 71% of consumers had used Google for product research in the past month and 26% had used ChatGPT. The second number is the one most startups have no segment for.

The traffic I want separated before I believe a conversion rate

Adobe's first-quarter 2026 analysis, covering more than 1 trillion visits to US retail sites and more than 5,000 survey respondents, found AI-referred traffic up 393% year over year.

What Adobe compared, March 2026	AI-referred traffic against other traffic
Conversion rate	42% higher
Revenue per visit	37% higher
Engagement rate	12% higher
Time on site	48% longer
Pages per visit	13% more

A year earlier, AI traffic converted 38% worse than human traffic. The reversal took 12 months. VentureBeat reports that LLM-referred traffic converts at 30 to 40% for companies actively optimizing for it. If analytics cannot tell an AI-referred visit from an ordinary organic visit, I am averaging a new channel into an old one and then deciding the new one does not exist.

The five numbers, and the order I demand them

Share of citation across engines. This is the brand's share of qualifying citations in the category's answers. Divide by the answers instead and the result is citation rate, a different number from the same sample. I track it on ChatGPT, Perplexity, Gemini, and Claude, because a brand can be cited in one and absent from another on the same query. A 2026 study of GEO visibility put citation inequality across engines at an average Gini coefficient of 0.715 (arXiv 2604.07585). A handful of domains take most of the citations. I want to know whether we are in that handful. On the monitoring set I have run, that is 35 queries, read daily. One engine is not the set.

Conversion of the referred visit. Perplexity sends referrer: perplexity.ai. ChatGPT sends traffic through OpenAI domains. Gemini can be isolated from ordinary Google organic with some work. Once the segment exists, I want conversion, revenue per visit, and depth. Adobe's 42% conversion premium and 37% revenue-per-visit premium are retail figures, not a forecast for a software pipeline. They are the reason the segment is worth building. If the analytics product cannot do it, server logs still have the referrer. The header was there. The report was not.

Stability across repeated runs. A single snapshot lies. The same study found that cited sources vary across repeated identical queries. Brand mentions are more stable than source URLs: brand-level Jaccard similarity of 0.46 to 0.48, against 0.35 to 0.40 for specific URLs. Stable is not the same as fixed. A brand in 7 of 10 identical runs is in a different position from a brand that appeared once. I aggregate runs. I do not brief the company on a Tuesday afternoon.

Earned citations against owned ones. Engines cite third-party publications far more often than they cite the brand's own site. I want the ratio on our query set. If the blog is cited never and a single contributed article is cited in three engines, the ratio is the investment memo. VentureBeat's line is the one I keep: the useful default is a citation map of where the model pulls from, how often the brand shows up, and how it is described. The owned-traffic report is a different map.

A path from the visit to the pipeline. This is the hardest of the five and the one I will not pretend we have on day one. Last-click and ordinary multi-touch models were built for paid and organic. An AI-referred visitor who then searches the brand and converts gets credited to Google. I start with branded-search lift. When citation frequency on a fixed set of 50 to 100 buyer queries rises, I look for the same movement in branded search, direct traffic, and demo requests. That correlation is the first proxy. It is a proxy. I label it as one.

The memo I would send back

I would return a budget note that gives me one run of one prompt, a blended "AI score," and a request to scale. The Gini coefficient of 0.715 says citations concentrate in a handful of domains. A single pleasant answer does not tell me whether we are in the handful. Brand-level Jaccard similarity of 0.46 to 0.48, and 0.35 to 0.40 for specific URLs, says the next identical run will often name someone else. I want the 35-query set, or whatever fixed set we can actually run, repeated. Seven appearances in ten runs is a position. One appearance is an anecdote.

I would also return a note that reports conversion without a segment. Adobe's retail cut, more than a trillion visits, showed AI-referred traffic up 393% year over year, converting 42% higher, with 37% higher revenue per visit, 12% higher engagement, 48% longer sessions, and 13% more pages, in March 2026, after a year in which the same traffic converted 38% worse. That reversal is why I am willing to build the segment. It is also why I will not paste 42% into a software forecast. The segment has to be ours: perplexity.ai as a referrer, OpenAI domains separated, Gemini pulled out of ordinary Google organic where we can. VentureBeat's 30 to 40% conversion band applies to companies already optimizing the referred visit. It is a ceiling I want to inspect, not a base case I will sign.

The proxy I will accept, while the attribution model still credits the follow-up Google search to Google, is branded-search lift against citation frequency on 50 to 100 buyer queries, with direct traffic and demo requests in the same weekly note. Forrester's matrix is why I label that note engagement value rather than a closed-won ROI. Productivity metrics move fast and are the wrong axis. Strategic value is slower. Pretending they share a dashboard is how the program looks inconsistent and then gets cut for the wrong reason.

Alphabet's $60.4 billion in search revenue, up 19%, and the 71% of consumers who used Google for product research in the prior month, keep the old dashboard alive for the old channel. The 26% who used ChatGPT are the line item the old dashboard cannot see. I fund the five metrics so that 26% has a number. I do not fund them so a score can be screenshotted.

The decision

I would approve a new AI-visibility spend only when three of the five are live. The query set exists and has been run more than once, so share of citation is a trend and not a screenshot. AI-referred visits are segmented, even roughly, so a conversion claim can be checked. And someone can say which pipeline number we expect to move, with branded search as an acceptable proxy if the direct line is not built yet. I would not approve the spend on a score inside a tool whose prompts, engines, and dates I cannot re-pull.

The scope is the budget decision for citation and answer-engine visibility. It does not judge the Adobe retail premiums as our premiums. A 393% traffic increase on US retail sites is a channel forming. It is not our pipeline. It does not say productivity metrics are useless for other AI projects. Forrester's split is why I keep those projects on their own timeline.

I would pull the spend back if, after a quarter on a fixed query set, share of citation did not move and neither did branded search, direct traffic, or demos. I would also refuse to scale a program whose "ROI" is a single run of a prompt, given Jaccard similarity for brands at 0.46 to 0.48 and for URLs at 0.35 to 0.40. One lucky answer is not a trend. I would change the gate itself if a later Adobe-scale or first-party cut showed AI-referred visits converting at or below other traffic for a full year after the 12-month reversal Adobe already recorded. That result would put the commercial argument back where it was when AI traffic converted 38% worse, and I would fund the work only for the strategic reason, with eyes open.

The stack I use so the five numbers stay attached to a real claim is Machine Relations. Share of citation, citation rate, and earned authority are defined in the glossary, and I want the meeting to use those definitions rather than a renamed share of voice. The publications have not been replaced. The reader that cites them has been added. The measurement has to include that reader or the budget is a guess with a dashboard in front of it.

FAQ

Which number do I require first?

A fixed query set, run more than once, reported as share of citation per engine. The Gini coefficient of 0.715 says the citations concentrate. A weekly screenshot from one engine will put us in or out of that concentration by accident. The set I have used is 35 queries, daily. A smaller company can start at weekly. It cannot start at once.

Why not judge this on rankings and organic clicks?

Rankings and clicks describe a list of links. An answer engine synthesizes a recommendation and may never send the click. Forrester's April 2026 piece calls the ROI failure a measurement failure. Alphabet's $60.4 billion of search revenue, up 19%, says the list still matters. The 26% of consumers who used ChatGPT for product research last month are simply not in that list's report.

How often is often enough?

Weekly at minimum, daily when the set is small enough to run. Brand-level Jaccard similarity of 0.46 to 0.48 means a repeated query agrees with itself less than half the time at the brand level, and URL agreement is lower, 0.35 to 0.40. I aggregate. A single afternoon is a sample of one.

What counts as pipeline evidence before the attribution model is rebuilt?

Branded-search lift, direct traffic, and demo requests, read against citation frequency on 50 to 100 buyer queries. I call that a proxy. Adobe's retail figures, 42% higher conversion and 37% higher revenue per visit, are the reason to build the real segment. They are not a number I paste into our forecast. VentureBeat's 30 to 40% conversion band is for companies already optimizing the referred visit. It is a ceiling to investigate, not a base case.

When do I stop the spend?

After a quarter of a fixed set with no movement in citation share and no movement in the proxy. Also immediately, if the only evidence in the budget memo is one prompt run on one engine. The 12-month reversal Adobe recorded, from 38% worse conversion to 42% better, is why I am willing to fund the segment at all. It is not why I fund a program that cannot show its own queries.