A founder asked me, fairly, whether any of this was worth it. He had done a little AI-search work, and his analytics showed maybe four visits a month from ChatGPT and a couple from Perplexity. Next to that number, the whole exercise looked like a rounding error. I looked at the same dashboard and agreed with him. As referral traffic, it was nothing. Then I asked him about the last three deals in his pipeline. In two of them, the buyer had walked into the first call already repeating something an AI engine had told them about his product, one flattering and one wrong. Neither of those buyers showed up in his analytics as coming from AI. They showed up as a booked demo, or as nothing at all until they surfaced in his CRM.
He was trying to price the influence by the traffic. The traffic was the smallest, most visible part of the return, and he was treating it as the whole thing.
The ROI question smuggles in an assumption that does not hold
“Is it worth it” sounds neutral. It is not. Buried inside it is a specific idea of how the return arrives: as measurable clicks, attributable to a source, that you can multiply by a conversion rate and a lifetime value and divide by what you spent. That formula is all over the search results on this exact question, and for paid ads or classic organic search it works, because the return really does arrive as traceable visits.
On AI search, most of the return does not arrive that way. A meaningful share of AI-driven referrals never gets attributed at all, because they leak into Direct traffic with no referrer to reveal where they came from. And the larger part never becomes a click in the first place. The engine answers the buyer’s question inside the chat, the buyer forms a view, and they arrive at your site or your sales call already carrying it. There is no visit to attribute because the influence happened before any visit. A click-based ROI does not undercount this return. It cannot see it.
Six times almost nothing is still almost nothing
The other thing the ROI posts lean on is the stat that AI-referred visitors convert several times better than other channels. It tends to be true, and for an early-stage company it is a trap. A high conversion rate on a tiny number of visitors is a big multiple on a small base, and the product of the two is still small. Quoting the multiple makes the work look impressive on a slide and changes nothing about the pipeline.
Worse, the conversion rate only measures the buyers who clicked through. It says nothing about the ones who ran their evaluation entirely in the chat and never visited at all, which for a lot of B2B categories is now most of them. You are computing a precise return on the sliver of buyers who behaved the old way, and ignoring the majority who did not. The precision is real. It is precision about the wrong slice.
The return you cannot put on an invoice
Here is what AI search actually moves for an early program, none of which has a row in an attribution report. Whether you make the shortlist the engine assembles before a human is involved. Which label you arrive with, the useful one or the diminishing one a competitor’s page wrote for you. How much of the sales call is spent selling versus correcting, because a buyer who was told the right things arrives ready to go deeper, and a buyer who was told the wrong things arrives skeptical or does not book at all. And the deals that quietly never happen, the buyer who got a wrong price or a wrong fact in the chat and crossed you off before you knew they existed.
That last one is the point. The most expensive outcome of ignoring AI search does not show up as a bad number. It shows up as an absence: pipeline that never formed, on evaluations you were never part of. You cannot bill for preventing it, and you cannot see it in the traffic, which is exactly why an ROI framework built on traffic will tell you it is not worth doing right up until it obviously was.
So is it worth it? That is the wrong question.
The useful question is not “what is the return.” It is “is the decision already happening here.” Those are different questions with different answers, and only the second one is answerable today.
You answer it by running the queries. Ask the engines the buying questions in your category, the way a prospect would, and watch what comes back. If the engine is already naming products, assigning roles, quoting prices, and building shortlists in your space, then the decision is already happening there, whether or not you spend a dollar on it. And that changes what the choice actually is. It is not “invest in AI search or not.” The exposure exists either way. The only choice is whether you author what the engine says about you or leave it to be authored by review sites, competitors, and stale pages. Doing nothing is not the zero-cost option. It is just the option whose cost never lands on a report.
How to judge an early program without a clean number
If you cannot price it by traffic yet, judge it by leading indicators instead. Are the retrieval engines crawling and citing your pages at all. Is the label they use for you correct and improving. Is branded search creeping up in the days after you show up in more answers, which is often the first visible trace of the influence that attribution misses. Those move weeks before any traffic number does, and they tell you the program is working long before the traffic would.
And expect the traffic itself to stay flat for a while, because it takes time to compound. If you demand a click-ROI that pencils out in month one, you will kill a program in its flat stretch that was about to bend. The number that would justify the work arrives after the work, not before it.
What I got wrong
I used to build clients an AI-search ROI model early, because it felt responsible. AI-attributed visits times conversion times deal size, tracked monthly, so we could prove the spend. The model was honest and it was useless. The numbers were tiny and jumped around, and for months they made the work look barely worth doing, which was a lie the math told with a straight face. What it left out was the entire off-click return: the buyers arriving to calls already correctly informed, the competitor’s mislabel that stopped sticking, the wrong price that stopped circulating. None of that had a referral row, so the model scored it as zero.
Now I do not lead with an ROI model for a young program. I lead with the exposure check, running the buyer queries live so the founder sees for himself what the engine already tells his buyers. And I say the uncomfortable part plainly: if you need a clean click-based ROI to greenlight this, the honest answer is that the number will not justify it yet, and the number is the wrong test. The right test is whether your buyers are deciding in a room you are currently not in.
Why this matters
The ROI question feels disciplined. It is the responsible-sounding way to defer the decision until the data justifies it. But it quietly assumes you can wait, that the cost of waiting is zero and the meter only starts when you choose to invest. On AI search the meter is already running. Every week you spend proving the ROI, the engines are handing your buyers a version of you that you did not write, and the buyers are believing it, and some of them are walking without ever becoming a number you could have measured.
You do not get to start the clock when the ROI clears. The exposure started without you. The only question the math cannot dodge is whether you want a say in what it says.
If you want to see what AI search currently tells your buyers, and whether the decision in your category has already moved into the chat, that is the first thing I check in an engagement, and the full method is on the methodology page.
If you want to measure the return, see the tools I actually use to track AI search visibility.