The Business Value of AI Recommendations Goes Beyond ChatGPT Referral Traffic

The business value of an AI recommendation extends past ChatGPT referral traffic, because a recommendation also moves brand search, direct visits and self-reported discovery that no referral report records. Two of the five studies reviewed for this article on September 23, 2026 measure that uncounted part directly. Referral traffic records one kind of journey: an AI answer shows a link and the buyer clicks it at once. Standard analytics then credits the last observable channel, so a shortlist formed inside an answer is filed later as organic search or direct traffic.

A Scrunch study published on June 17, 2026 gives the gap a consumer-market signal, an Omniscient Digital analysis published on August 28, 2026 gives it a B2B one and the Canlah AI self-audit of September 7, 2026 supplies a smaller first-party example. For brands selling in Singapore and across APAC, where buyers ask in English and Chinese across several engines and long sales cycles hide the first touch, the gap is wider. Referral traffic is the floor of AI recommendation value, not the measure of it.

Key findings

• Recommendations move later behaviour: Scrunch reported on June 17, 2026 that new users recommended a brand by an AI assistant searched it on Google at 9.24% the following week, against a 3.28% matched baseline.
• Click-based attribution misses most AI-sourced leads: Omniscient Digital found that first-touch attribution credited 28 of 189 leads who said an AI tool sent them, about 15%, in its August 28, 2026 analysis.
• Canlah AI cannot yet price its own AI visibility: its September 7, 2026 self-audit recorded 7 non-brand mentions in 177 grounded answers, but its tracking probe found no conversion event to connect those answers to sign-ups.
• A four-layer measurement model: answer exposure, assisted demand, first-party attribution and commercial outcome should be recorded separately and joined without counting the same conversion twice.

What the referral and recommendation studies measured

Five studies published between June 16, 2025 and August 28, 2026 measure AI recommendation value in four different units, so their figures should not be averaged. Two of them measure what happens when there is no click at all, and the other three measure what a visitor who does click is worth.

Five AI referral and recommendation studies reviewed September 23, 2026.

Study	Sample and window	Reported result	What a buyer should do with it
Scrunch, June 17, 2026	Opt-in panel linking AI conversations to the same users’ web activity, February to May 2026	After a recommendation, brand search rose from 3.28% to 9.24% and site visits from 3.43% to 7.45% within a week	Budget for brand search and direct visits, not referral clicks, as the carrier of the effect.
Omniscient Digital, August 28, 2026	213 inbound leads with a free-text discovery answer, two years of CRM data	189 named an AI tool; first-touch credited 28 to AI referrals and filed 95 as organic search and 59 as direct	Add a free-text discovery field before judging AI search, or see about 15% of it.
Ahrefs, June 16, 2025	Ahrefs’ own site, 30 days	AI search sent 0.5% of visitors and 12.1% of sign-ups	Judge the referral stream by conversions, not by its share of sessions.
Semrush, June 2025	Semrush AI search traffic study; sample size and window not found on reviewed page	The average AI search visitor was 4.4 times as valuable as an organic visitor, by conversion rate	Reproduce the multiple on your own analytics before using it as a planning input.
Search Engine Land report, October 23, 2025	Ecommerce sites comparing ChatGPT referrals with other channels; site count not found on reviewed page	ChatGPT referrals were about 0.2% of sessions; organic search converted about 13% better and affiliate about 86% better	Do not assume a per-visit premium in ecommerce; measure it by category and intent.

Figures were read from the publishers’ pages on September 23, 2026 and were not independently re-run. “Not found on reviewed page” means the publisher did not state that detail on the page reviewed on that date, and is not proof that the detail does not exist.

Sources: Scrunch prompt-to-purchase study; Omniscient Digital attribution research; Ahrefs AI search conversions; Semrush AI search traffic study; Search Engine Land conversion report

Four of the five publishers state a sample and a window on the reviewed page, and Scrunch and Omniscient Digital also publish the matched baseline or the raw lead counts behind their headline numbers, which is why those two carry the argument here. They also sell GEO software, SEO software or agency services, so the findings are vendor research rather than neutral consensus. The Scrunch panel was more AI-engaged than an average buyer, its replicated categories were beauty, apparel and audio, and Omniscient Digital measured one agency’s own pipeline.

The disagreement in the last three rows is informative: per-visit conversion rates depend on the site, the offer and the query mix, while the first two rows do not depend on referral clicks at all.

Scrunch also separated passing mentions from active recommendations. A mention raised the probability of a brand search by 3.3 percentage points and a site visit by 1.9 points; a recommendation raised them by about 6 and 4 points. The study supports a distinction between being named and being put forward, but supplies no universal conversion rate.

Sources: Scrunch prompt-to-purchase study

Why ChatGPT referral traffic misses part of the journey

ChatGPT referral traffic loses the journey at four points: the answer that resolves a shortlist without a click, the branded search that follows days later, the third-party page that carries 87% of the citations and the conversion event that was never configured. Each point is a separate measurement fix, and only the last one is under the brand’s own control.

Most influence produces no click

An answer can resolve a shortlist without sending anyone anywhere. Omniscient Digital reports that its most-retrieved page appeared in AI answers 210 times in one month and drew 10 referral sessions.

The first visible touch is usually not the AI one

A buyer who sees a recommendation in ChatGPT and later searches the company name is credited to organic search, and one who types the domain is credited to direct. First-touch and last-touch models record the earliest or latest tracked session, and the AI answer was neither.

Much of the evidence sits on other domains

Omniscient Digital reports that 87% of the AI citations naming it sit on third-party pages, based on Peec AI data. Listicles, reviews and comparisons shape the answer, and no first-party or server-side tracking can observe a buyer reading them inside an assistant.

Measurement ends before the commercial result

A referral session can be counted without a configured conversion event, and a visibility gain can be reported without a lead. The business value of AI recommendations can only be tested when the answer record, the session record and the CRM record share identifiers and dates.

A first-party view from Canlah AI

Canlah AI audited its own site, canlah.ai, on September 7, 2026, using the same pipeline it runs for client diagnostics. Thirty-nine buyer questions were sent three times each to ChatGPT and Gemini, and 216 answers had grounding sources. Canlah AI was mentioned in 7 of 177 grounded answers to open buyer questions, a 4.0% rate, and in all 39 grounded answers to questions that named Canlah.

The branded result matters for this argument. Four of the branded questions asked whether Canlah is a legitimate company in Singapore, both engines named Canlah AI in all 23 grounded runs, and the most-cited sources in that group were registry and directory pages such as sgpbusiness.com and companieshouse.sg. That verification step happens inside the answer, so a buyer who is reassured and later books a call leaves no referral session behind.

The same audit exposed the measurement gap. The tracking probe detected Consent Mode v2 on the homepage but no conversion event, and it reported no GA4 tag although one was visible in the page source on September 23, 2026, so its zeros are treated as unverified. Canlah AI has not published a referral-to-sign-up figure for its own site, which is a gap in Canlah AI’s own process rather than a client result.

The evidence shows that answers, verification and sessions are different events. It does not show how much revenue any of them produced.

Mention, recommendation and citation are not interchangeable

Mention, recommendation and citation describe three different answer states, and only one of the three is a commercial endorsement. A citation means a page supplied information, and the brand may never appear in the prose. A mention means the entity was named, but the answer may be neutral or negative. A recommendation means the assistant presents the brand as suitable for the buyer’s need.

Position and framing set prominence, while factual accuracy decides whether the exposure is safe. In the Canlah AI self-audit, ChatGPT opened canlah.ai during its web search in four of its seven non-brand mentions, and in every run where it opened the page it went on to name the brand. A citation-only score would have missed the runs where the page was never retrieved, and a referral-only report would have missed all seven.

A four-layer model for measuring AI visibility ROI

Four layers separate what can be observed from what can only be declared: answer exposure, assisted demand, first-party attribution and commercial outcome. Layers one and two can be collected without the buyer’s cooperation, while layers three and four cannot exist without form fields and CRM access.

The four layers of an AI visibility ROI record, as used in Canlah AI reporting on September 23, 2026.

Layer	What to record	Decision it supports
1 Answer exposure	Buyer question, engine, language, market, run count, mention, recommendation, position, cited URLs and the full answer.	Whether the brand is present, accurately described and actively put forward.
2 Assisted demand	Brand-query trend, direct visits, returning users and regional changes, dated against answer changes.	Whether demand patterns follow AI exposure, without claiming attribution.
3 First-party attribution	Free-text discovery answers, demo notes, sign-up fields, CRM source detail and the question the buyer says they asked.	Which people or accounts declare AI-assisted discovery.
4 Commercial outcome	Qualified lead, pipeline stage, revenue, retention and disqualification reason, joined by a stable account identifier.	Whether AI-influenced discovery reaches a business result.

Layers one and two come from probe records and analytics owned by the brand. Layers three and four depend on client systems, so a report that covers only layers one and two should say so rather than describe the gap as a null result.

Two labels keep the layers honest. AI-referred marks sessions with a detectable assistant referrer. AI-influenced marks journeys supported by self-report, sales notes or another approved first-party signal. A conversion can carry both labels but must be counted once.

What changes for Singapore and APAC buyers

B2B buying cycles in Singapore run longer than the one-week consumer journeys the Scrunch panel measured in beauty, apparel and audio. A recommendation may shape a shortlist weeks before procurement, legal or technical review, so the later signal is often a documentation visit, a partner introduction or an enterprise demo. Attribution has to be account-aware and tied to CRM stages rather than to sessions.

APAC also requires measurement by language and engine. A ChatGPT click pattern cannot be assumed for DeepSeek, Qwen or Doubao, and a Chinese-language recommendation may lead to a local search engine, a WeChat conversation or an offline referral that no Western analytics stack records. The Canlah AI self-audit carries the same blind spot: it sampled ChatGPT and Gemini only, so it holds no evidence about how DeepSeek, Qwen or Doubao answer the same questions.

What buyers should test

• Freeze 20 to 40 non-branded buyer questions, sample each several times per engine and report mention and recommendation separately from citation.
• Configure the conversion events that matter before any GEO work starts, not after the first report.
• Add a free-text “how did you hear about us” field to demo and contact forms, and classify answers that name ChatGPT, Perplexity, Gemini or another assistant.
• Record assistant referrers as a separate channel group, and keep AI-referred and AI-influenced labels distinct.
• Join answer records to CRM records by date window and account, and count each commercial outcome once.
• Label before-and-after results as association unless a control design supports causation.

Where Canlah AI fits

Canlah AI is a Singapore-based SEO + GEO agency that measures and improves brand visibility inside ChatGPT, Gemini, Google AI Overviews and Google AI Mode, reporting AI visibility as re-verifiable ranges with timestamped evidence. Its service pages describe a monthly re-probe of a locked buyer-question pool, a citation-frequency comparison against baseline and AI-referral attribution from the client’s own analytics.

It is most relevant to B2B and export-focused brands that need the answer layer measured and joined to their own CRM, not to every company seeking a low-cost referral dashboard. Its own data shows the current limit: monthly reporting covers answer exposure and assistant-referred sessions, while self-reported discovery and pipeline joins depend on client form fields and CRM access, and Canlah AI has not yet closed that loop on its own site.

Methodology and limitations

Third-party figures come from public pages reviewed on September 23, 2026. They are vendor-published research, not an independent B2B benchmark, and they differ in unit, window and market. Canlah AI figures come from its September 7, 2026 audit of canlah.ai, one property in one category, sampled on ChatGPT and Gemini with three runs per question. The data has no control group, no conversion events and no referral sessions, so it identifies measurement gaps, not incremental revenue caused by AI recommendations. Every third-party figure above is linked to its source page, and buyers should re-check each one on that page before it enters a business case.

Frequently asked questions

What is the business value of an AI recommendation beyond ChatGPT referral traffic?

The measured part is later behaviour. Scrunch reported on June 17, 2026 that brand search after a recommendation ran at 9.24% against a 3.28% matched baseline, and site visits at 7.45% against 3.43%. Omniscient Digital reported on August 28, 2026 that self-reported discovery found 6.8 times as many AI-sourced leads as first-touch attribution, so referral sessions record the clicks and not the shortlisting inside the answer.

Can AI search ROI be attributed precisely?

No. Direct AI referrals can be observed precisely, but assisted influence needs self-reported discovery, CRM evidence, brand-demand trends and, where practical, controlled tests. A credible AI visibility ROI report separates observed referrals, supported influence and modelled incrementality.

How do teams avoid double counting in AI search attribution?

Assign a stable lead or account identifier, keep the first and latest known touchpoints and define a lookback window. Count each commercial outcome once, then attach AI-referred and AI-influenced labels as supporting evidence. Direct, organic and self-reported AI conversions should never be added together as separate customers.

Related articles

• The Own-Brand GEO Audit Experiment
• How to check if ChatGPT recommends your brand
• Google AI Overviews are changing SEO
• GEO vendor evidence checklist