Does AI Search Traffic Convert?

The direct answer: 12% of the traffic produced 75% of the commercial intent

GrantCompass measured 3,392 sessions across three usable acquisition channels between June 1 and August 24, 2026. AI assistants — ChatGPT, Microsoft Copilot, Perplexity and Claude, identified by referrer — supplied 417 of those sessions, or 12.3%. Organic search supplied 1,149. Over the same window the site recorded four real checkout starts on its $9 product and one completed sale. Three of the four checkout starts, and the single sale, arrived from an AI assistant.

Expressed as a rate against each channel's own sessions, AI-referred visitors started checkout at 0.72% and organic-search visitors at 0.087% — a ratio of roughly 8 to 1. The sample is four checkout starts and one sale, and no amount of framing makes that a large number. The asymmetry is the finding worth publishing: a channel carrying an eighth of the traffic carried three quarters of the purchase intent, in a market where the received wisdom is that AI assistants answer the question and keep the click.

Sessions, engagement and depth by channel

Three channels carried measurable volume in the window. Organic search and AI assistants are clean measurements; Direct is not, and is shown with that caveat attached rather than dropped, because Direct is where the missing AI traffic hides.

Two comparisons matter and they point in opposite directions. AI-referred visitors go 22% deeper — 2.30 pages per session against 1.89 — while spending 27% less time per session (79.7 seconds against 108.9). On the standard engagement metric the two channels tie: 54.9% of AI sessions were engaged against 57.4% of organic sessions. An AI referral is therefore not a "better-engaged" visitor by GA4's definition. An AI referral is a faster one, sampling more pages in less time, and on this site that behavior converted eight times more often. Direct's profile — 1,826 sessions at a 20.4% engaged rate and 1.47 pages per session — is the signature of a channel diluted by automated traffic.

An AI referral arrives mid-decision, not at the start of research

An AI-referred visitor lands on grantcompass.co already knowing what to look for, because the assistant has already narrowed the question. The channel average is 2.30 pages per session, but the average conceals the shape: engaged AI sessions browse between 5 and 30 individual program pages in a single visit, and one visitor read a single program page for 4.5 minutes — more than three times the 79.7-second channel average for an entire session. Shallow bounces and deep sweeps, with little in between.

The single completed sale is the clearest record of the behavior, step by step:

1. ChatGPT recommends GrantCompass in answer to a funding question.
2. The visitor lands on /small-business-grants-open-now, a dated, verified "what is open right now" status page.
3. The visitor clicks straight into the Closing soon section.
4. The visitor opens an individual program page.
5. The visitor starts the $9 checkout and pays — 61 seconds from landing.

AI assistants cite dated status pages, not evergreen guides

Five pages received the AI-referred landings in this window, and the pattern across them is the most portable finding in the study. Ranked by the traffic they took, the landing pages were /small-business-grants-open-now, /corporate-small-business-grants, /womens-business-grant-calendar, /women-owned-business-grants and /grants-for-women-of-color. Four of the five are dated status or calendar surfaces that answer "what is open, and by when". None is a general explainer.

The shape being rewarded is a verified, dated, current-state answer — the class of fact a language model cannot produce from its own weights and must therefore fetch and attribute. Evergreen how-to guides on the same domain, covering the same topic with more words and more internal links, took no measurable AI referrals over the same 85 days. Timing follows the same logic: the first citation of a newly published status page appeared roughly 24 to 48 hours after the page was pushed into Bing's index by direct content submission, not weeks later.

AI referral volume grew 6.8× from June to July, then flattened

Monthly volume is reported by assistant, because the mix moved as fast as the total. August covers 24 days, not a full month, so the daily rate is the honest comparison: 1.2 sessions per day in June, 8.1 in July, 5.4 in August. July was the step change; August is running roughly a third below July's pace, which is a plateau rather than the compounding curve a single month of growth would suggest.

ChatGPT accounted for 246 of the 417 AI sessions (59%) and Microsoft Copilot for 136 (33%) — together 92% of the channel. Perplexity peaked at 21 sessions in July and fell to 4 in the first 24 days of August. Claude moved the other way, from 0 in June to 7 in August. For a site of this size the operational reading is narrow: two assistants are the channel, and both are fed by Bing's index.

Every AI number on this page is a floor

Referrer stripping is the central measurement problem in AI-search analytics. Assistants deliver links inside app webviews, through redirect wrappers and from clients that send no referrer at all, and an industry-measured 70–80% of AI-referred traffic consequently reaches analytics with no attribution. Every AI figure in this study is derived from the minority that keeps its referrer, so 417 sessions is a lower bound on the real total and 12% is a lower bound on the real share. GrantCompass does not publish a corrected estimate, because a correction inferred from someone else's strip rate is not a first-party measurement.

Direct is the second half of the same problem. Direct carried 1,826 sessions at a 20.4% engaged rate and 1.47 pages per session, and it contains at least three populations that cannot be separated after the fact: genuine returning visitors, documented scraper waves on August 12 and August 15, and stripped AI referrals. The engagement profile is depressed enough to prove that automated traffic is present, and no post-hoc rule reliably splits a stripped ChatGPT click from a bot hit.

What this study does not establish

Sample size. Four checkout starts and one completed sale. A single additional organic-search purchase would move the ~8× ratio materially, and any confidence interval around that ratio would be wide enough to include much smaller effects. The asymmetry — 12% of sessions producing 75% of starts — is the claim; the multiple is an illustration of it, not a benchmark.

Scope. One website, one vertical (US small business funding), one price point ($9), one 85-day window, one product that existed for only 5 of those 12 weeks. Grant discovery is unusually deadline-driven and unusually well-suited to assistant-mediated research, so the result should not be read as a general e-commerce finding.

Attribution. No holdout, no randomization, no controlled experiment. Channel assignment is last-non-direct referrer, which credits the AI assistant for a visitor who may have first encountered the brand elsewhere. Cross-device and multi-touch journeys are invisible. GrantCompass reports an observed association between channel and purchase, not a causal effect of AI citation on revenue.

Methodology, window and exclusions

Window. June 1 to August 24, 2026 — 85 days, the site's first 12 weeks with measurable AI-referral traffic. The funnel went live June 20, 2026 and the $9 product launched July 21, 2026, so the four checkout starts fall in the product's first 5 weeks while the session denominators span the full 85 days. Both channel rates are understated in the same direction by that mismatch. The page title rounds 85 days to 90.

Channel definitions. Google Analytics 4 sessions, grouped by referrer. AI assistants = referrals from ChatGPT, Microsoft Copilot, Perplexity and Claude. Organic search = GA4's organic search channel. Direct = sessions with no referrer, reported with its bot caveat rather than excluded. Engaged rate is GA4 engagedSessions / sessions. Bot waves were flagged on Direct only; the AI and organic channels showed no bot contamination in this window.

Conversion. Checkout starts and completed payments come from Stripe on the live GrantCompass US account. The operator's own test purchases are excluded from every conversion count by a Stripe-side rule — the four checkout starts are real visitors, which is why the number is four and not larger. Referrer attribution for a checkout start is the session's last non-direct referrer. Ground truth. Every figure on this page is frozen in docs/seo-sprint-2026-08-23/data-ai-conversion-study.json, generated August 24, 2026; no number here is computed anywhere else.