AI Search ROI Can't Rest on 'How Did You Hear About Us' Answers

Your 2027 budget case probably needs a line for AI search, and finance will want a return next to it. On October 6 Ahrefs published a method that works out the quarter's return two ways: from conversions your analytics can tie to visits from AI tools, and from what new customers say when asked how they heard about you. In its worked example the first comes out at 33 percent and the second at 100 percent, and the advice is to "lead with the higher figure" (Ahrefs).

Ahrefs is candid that neither number is exact, and it expects both to undersell AI. The survey number has a different problem: it answers a question finance is not asking. It can show how much of your revenue passed through an AI tool on the way in, but not what the AI work on your budget line earned. The gap between those two opens in three places: how the answers are credited, how the question is asked, and what the formula leaves off the revenue side. Dealing with all three leaves you a smaller plan that is much harder to knock down.

Each AI mention gets credit for the whole sale

The survey half of the method lets each new customer pick more than one answer, adds up the revenue from everyone who picked an AI tool, and applies that share to all of the quarter's new revenue (Ahrefs). Picture a $9,000 customer who ticks Google, ChatGPT and a podcast. All $9,000 goes into the AI total, and if you run the same sum for Google, the same $9,000 lands there too, so the channel totals add up to more money than you booked. The method never splits the sale.

The attribution guide in marketingskills, an open library of instruction files that marketers load into AI assistants, warns about exactly this step: discount survey share before you turn it into credit. Its example is that 22 percent of customers recalling YouTube "is evidence of awareness, not a claim that YouTube deserves 22% of conversions" (marketingskills).

Ahrefs' own headline figure is almost exactly that case: 21.7 percent of its signups name ChatGPT or Claude as how they found the company (Ahrefs). That tells Ahrefs how often AI turns up in its buyers' paths, which is worth knowing. It does not tell anyone what share of the revenue AI produced.

The question decides how often the answer is AI

How often AI turns up also depends on the form, and Ahrefs recommends changing it by adding AI tools to the options in your "How did you hear about us?" drop-down (Ahrefs). That is sensible, and it also swaps the measuring instrument partway through the year.

Offering an answer makes people more likely to pick it. In a Pew Research Center poll after the 2008 election, 58 percent chose the economy as their deciding issue when it appeared on a list, while only 35 percent named it when answering in their own words, and every other issue also did better when offered (Pew Research Center). SparkToro founder Rand Fishkin points to a bias specific to this moment: whatever channel is "hot right now" gets overweighted in these surveys, and he names AI answers as today's example (Rand Fishkin).

So a team that adds ChatGPT to its drop-down in October should expect AI's share to climb from the new form alone, and a slide comparing the fourth quarter with the third will present the form change as growth.

The revenue side never asks what you would lose

Fix the crediting and the question, and one gap remains that no survey can close. Ahrefs builds the cost side around a single test: "would you still pay for it if you stopped doing AI search work tomorrow?" Costs that would disappear count in full, and shared costs count in part (Ahrefs). The revenue side never gets asked the matching question: would this revenue still arrive if the AI work stopped tomorrow?

A lot of it would. By Ahrefs' own account, AI models learn what to say about a brand "largely from what the rest of the web says about you" (Ahrefs). Pages you do not own feed the answer, and Google's AI summary for a big brand's name cites Wikipedia more than any other source. A customer who met you in a ChatGPT answer met the result of all of that, and only part of it came out of this year's AI budget.

We made the same distinction about ChatGPT's ad dashboard in September: a conversion that follows an ad is not evidence that the ad caused it (Promptafire). The organic version is easier to miss, because the number comes from your own customers and feels more trustworthy than a platform's report.

What to put in the 2027 plan instead

None of this argues for leaving the AI line blank. The survey is still your best view of AI's footprint, meaning how much of your business passes through an AI tool on the way in, so keep running it and change what you call it. The marketingskills guide notes that an open question, "How did you first hear about us?", catches what a pick-list misses, and that a pick-list tilts the answer (marketingskills). Ask it open and sort the answers into channels afterwards. If you need a list, keep its wording identical from quarter to quarter, which is Pew's rule for measuring change over time (Pew Research Center). Treat the quarter you add AI tools as a break in the series rather than as growth.

Put tracked AI conversions beside it as the part of that footprint you can see, and mark any jump that comes from measurement. Since the end of September Gemini has tagged some of its links so analytics can see where a visit came from, and Google has not said which ones (Search Engine Journal). Visits that used to land in Direct will start showing up as Gemini and look like fourth-quarter growth.

Neither line is a return, and walking in without one costs something. Growth Memo warns that directors without a defensible case for AI search will lose that budget to paid media (Growth Memo). A slide with no 100 percent on it can lose the room to one that has it, until someone asks what the 100 percent would do if the AI work stopped. So fund the AI-specific work the way Growth Memo suggests treating search surfaces you have not tested: as learning work, with a capped budget, one owner and a kill date (Growth Memo). The kill date is the day the work stops unless a test has shown results.

AI answers give you no way to split traffic 50-50, so the workable test uses a control group: similar pages you leave alone, tracked in the same monthly log of which prompts cite you on ChatGPT, Perplexity and Google's AI answers as the pages you change. In a webinar with Search Engine Journal, the SEO platform seoClarity described a client test in which adding FAQ sections to a set of pages raised how often AI answers cited them against the control group, and removing the sections brought citations back down (Search Engine Journal). That shows what the work changed, which a survey never can. Take the test design to the budget meeting with its kill date already filled in.

Sources

• AI Search ROI Is Messy. Here's How to Measure It Anyway, Ahrefs, 6 October 2026
• How to Measure AI Search Visibility When Attribution Falls Short, Ahrefs, 6 October 2026
• How to plan 2027 when traffic is down and AI won't attribute, Growth Memo, 5 October 2026
• Attribution skill, coreyhaines31/marketingskills
• Writing Survey Questions, Pew Research Center
• Rand Fishkin on "How did you hear about us?" surveys, LinkedIn, 28 September 2026
• Google Gemini Adds UTM Parameters For Referral Attribution, Search Engine Journal, 1 October 2026
• AI Search is Working. How to Prove It With Real Tests., Search Engine Journal, 23 July 2026