Do AI Recommendations Generate Leads? The US Measurement Design, and the Figure Nobody Should Hand You Yet
Four paths run from an answer to your pipeline, three of them leave a record you can build, and this is the design that counts them with their denominators printed.
Published by AI Knows Us · Updated 16 October 2026
An assistant naming your company can produce a lead, and the part worth arguing about is which of those leads you can observe. Four paths run from an answer to a United States pipeline: a click on a credited link inside the answer, a search for your name afterwards, a name passed to a colleague, and your material used in an answer that credits somebody else. The first leaves a referrer, the second shows up as branded search and direct sessions moving together, the third arrives as a lead that says it heard about you somewhere, and the fourth leaves nothing. We publish no conversion rate and no lead count for this, because we have not run that study. What follows is the design, written so you can run it on your own pipeline this quarter.
The four paths, and what each one leaves behind
| Path | What it leaves in your data | Where you count it |
|---|---|---|
| Clicked a credited link in the answer | A session with a referrer, or sometimes none | Analytics, as its own channel group |
| Searched your name afterwards | Branded impressions and clicks, and direct sessions | Search Console and analytics, as a pair |
| Was told about you by somebody who asked | A lead with no digital trail | The form question and the sales note |
| Your page was used without credit | Nothing | Not countable, and worth saying so |
Three of the four are observable with work you already know. The fourth is real, because an answer can be built partly from your material and credit a different domain, so name it in the report as an uncounted path rather than letting a reader assume the total is complete.
Build the channel group from your own referrers, not from a published list
The temptation is to paste in a list of hostnames somebody published and call it AI traffic. Build it from your own data instead, because the values appearing in your reports are the values you have to group.
- Pull the full referrer or source and medium report for a long window, unsampled, and sort by sessions rather than filtering first. The candidates are visible by inspection and they change.
- Write the grouping rule as a dated file listing every string you included and every near-match you excluded. That file is part of the measurement, exactly like a question set.
- Keep a raw column and a grouped column, and never overwrite the raw one, because next quarter you will want to re-group this one under a corrected rule.
- Count sessions with no referrer as their own line. Folding them into direct hides the path that produces exactly that signature.
One warning about the arithmetic: a referral count is a floor on this rather than a measure of it. It counts clicks you received and says nothing about answers where you were named without a link, which is a separate audit with its own method, set out in a citation and link audit.
The form question that carries most of the signal
One question on the enquiry form does more work than any tagging scheme, and it has to be designed rather than bolted on.
- Make it a required field with a short option list and a free text box. Options a United States buyer will recognise, including one that names assistants, plus referral, search, social, event and other.
- Record the completion rate as a denominator. If 240 of 400 leads answered it, every figure from that field is out of 240 and the report says so. A self-reported source published without its completion rate is the softest number in this subject.
- Keep the free text and read it monthly. People describe this path in their own words long before an option list catches up.
- Stamp it at creation and never overwrite it. A field updated by whoever touched the record last describes your sales process rather than the buyer.
Reconcile with the CRM, and judge quality separately from volume
Stamp lead source at creation, carry it through every stage, and report four counts over the same window: leads, qualified leads, opportunities and closed business, each split by path with its denominator printed. The comparison that settles the argument is whether those leads qualify at a similar rate to your other paths, because high volume with poor qualification is a different decision from small volume with strong qualification. Picking one source of truth per question is worked through in attribution explained honestly.
The study we have not run, and the baseline we do publish
What we do not hold: any lead count, any click-through figure from inside an answer, and any conversion rate for this path, for our own business or anybody else. We have not run a lead attribution study, so we publish no number for it.
What we do hold is the answer side, measured twice. On 16 October 2026 we asked 210 buying questions per market on two engines, ChatGPT and Gemini, producing 1,260 readings with none discarded, and our own domain was credited as a source in 6 of them, which is 0.5%. The day before, a narrower file of 102 per market, designed eleven days earlier, left us in 3 of its 597 usable readings. Different questions, twice the sample, one answer, which is replication rather than growth. Add both runs together and Gemini used this domain in none of 921 readings. Every one of those counts is about being credited, and not one of them is about a lead.
When you run yours, publish seven lines: the window with both dates, the number of businesses, sessions by path, leads, qualified leads, the grouping rule version, and the self-report completion rate. Name the engines individually. Two engines sat in the runs above. Perplexity, Claude and Google AI Overviews did not, and all five are surfaces we read for customers, so a count from two of them is never described here as a count from five.
Before you measure the lead, check the document that would have produced it
A measurement study assumes the answer could have sent somebody. Sometimes the document stops it first, and the check is cheap.
Our own pricing page is the example. Fetched as GPTBot on 16 October 2026 it returned 88KB, of which 55KB was script payload, and the visible first response carried Rs 499, Rs 2,949, Rs 11,799 and Rs 35,399 and no dollar figure at all, because the currency conversion ran in the browser. An assistant asked what this costs in the United States was reading that document. The cause was not a broken detector: a browser sends a country header, a time zone or a language and gets its own ladder, a crawler sends none of them, and the fallback was deliberate. The defect was that the other currencies were never in the HTML. It was fixed the same day and the page now carries 15 dollar figures, 5 pound, 5 euro and 31 rupee in server rendered text on every request, and rendering and the first response an AI crawler reads covers the mechanism.
Run that check on your own pricing and contact pages before building attribution. A dollar figure that exists only after a script runs is a lead that was never going to arrive.
Questions about leads from AI recommendations
Why can I not see traffic from an assistant in my analytics?
Three ordinary reasons account for most of it. Some answers name a business without a link, so there is nothing to click. Some clicks arrive with no referrer and land in direct. And a buyer who reads an answer often searches your name later, which arrives as branded search. Build the no-referrer line and the branded search line into the report and the picture fills in.
Can an answer influence a sale with no click at all?
Yes, and the form question is how you catch it. A buyer given three names who then searches for one of them leaves a branded search trail and a direct session rather than a referral, which is why branded search volume and direct sessions belong in this report as a pair, read month over month.
What conversion rate should I expect here?
Measure your own, because a borrowed rate from a different offer, market and price point misleads in both directions. Any published industry figure here is invented until somebody shows the window, the sample, the engines and the denominators. Those four are what to ask for.
How long should the observation window be?
Ninety days, with the dates printed, and the same window repeated rather than a rolling comparison. Shorter windows in a United States business-to-business pipeline finish before the deals do.
The thing to settle before anybody sells you an attribution number
Every figure here is a fraction, and a fraction is only as good as the line saying what was counted and over what. A supplier who will tell you how many leads an assistant sent, but not the grouping rule, the completion rate or the window, has given you a presentation. AI Knows Us sells the answer-side measurement, the question set, the readings and the credited sources, and its honest boundary is that it tells you whether you were named and by whom rather than how many leads followed. Our own credited figure is 6 readings of 1,260 on 16 October 2026, published with its denominator in what AI assistants actually cite across 597 readings. Start with the form question this week. One field, no tooling, and the most reliable column in the report for a long time.