EVIDENCE + FRAMEWORK

Winning customers from ChatGPT and AI

We separate what ClassyDesign has actually measured, what official sources actually say, and what's genuinely controllable.

Last reviewed: September 29, 2026

Four different things, collapsed into one word

“Showing up in ChatGPT”, “getting traffic from ChatGPT”, “generating leads from ChatGPT” and “winning customers from ChatGPT” are not interchangeable.

AI Visibility
AI Referral Traffic
AI Lead Generation
AI Customer Acquisition

AI Visibility

AI surfaces or cites you in an answer.

AI Referral Traffic

A real visitor arrives at your site from an AI system.

AI Lead Generation

That visitor fills a form or makes contact.

AI Customer Acquisition

That lead becomes a paying customer.

Can AI systems really send traffic?

A visit is not a customer: visibility, visits, enquiries and customers are four different things (see the four layers above). On Classy's own site (a service business, not a shop), ChatGPT sent 4 of 39 external visits in 28 days (26 Aug–22 Sep 2026), and one of those visits included a completed contact form. That is a real but very small signal; whether ChatGPT caused any enquiry has not been verified. Four visits are too few for a conversion rate or a comparison with Google, and one business says nothing about other service businesses. We measure AI referrals and enquiries on a client's site by combining the observed referrer with a 'where did you find us?' question.

Verify these claims: the case page lists each one with the URL to check and a prompt you can hand to any AI assistant. Claims and where to check them →

How does traffic become a lead?

A visitor from AI needs the same conversion architecture as any visitor: a clear offer, visible trust signals, and a single-step contact path. The difference: an AI-referred visitor has usually already asked a question and received an answer; when they land on your site, they're checking that answer. If your page doesn't seem to recognise that question, trust disappears immediately.

INDEPENDENT EVIDENCE

What is known beyond our own measurement?

We cite two sources: a peer-reviewed study and an industry study that publishes its method and limits. Both say the same thing: AI-referred visits are real, growing fast from a tiny base, and still a small share of total traffic.

Kaiser & Schulze, Marketing Science (2026), peer-reviewed

973 e-commerce websites, first-party Google Analytics data, August 2024–July 2025; more than 50,000 ChatGPT-referred transactions against 164 million from traditional channels.

•ChatGPT-referred visits are under 0.2% of all visits.
•Conversion is below every channel except paid social: organic search converts about 13% better than ChatGPT traffic, affiliate about 86% better.
•Conversion rate rose steadily over the year; average order value fell.
•In complex product categories it takes a 4.6× higher traffic share and out-converts organic, direct, email, referral and paid social there.

E-commerce was measured, not service firms; carrying the complex-product finding over to services is an inference, not a measurement. We took the figures from the authors' summary site and press coverage; the full text was not accessible.

DOI 10.1287/mksc.2025.0489 ↗

SE Ranking, AI traffic study (June 2026), industry data

101,574 websites, aggregated Google Analytics data, January 2025–April 2026. A vendor source; its method and limits are published, and it does not measure conversion.

•Visits from AI search engines are 0.32% of total traffic (0.02% in 2024).
•74.8% of AI referrals come from ChatGPT, 11.6% from Gemini, 2.6% from Claude.
•AI visitors stay longer on average; the median gap is small (2:24 against 1:53) and the mean is inflated by outliers.

No industry breakdown; the "16× growth" headline comes off a 0.02% base.

seranking.com/blog/ai-traffic-research-study ↗

Our own figures should be read next to these benchmarks: on classydesign.studio, 4 of 39 external visits in 28 days came from ChatGPT; on turkdizisi.com, from 6 to 24 September 2026, about 52% of active users measured in GA4 first arrived via ChatGPT referral. The second is an outlier against these benchmarks: a content site, one period, not a revenue claim.

None of these sources says "AI visibility brings customers"; the first measures the opposite on aggregate. Neither do we. We do not cite vendor blogs whose method is not published.

Sources read on 26 September 2026.

What a business can do (effect on AI recommendations not measured)

There is no measured recipe for getting recommended; what a business can do is keep its information readable and verifiable, and measure what AI assistants actually cite and send. What can be measured is which sources AI assistants cite and which visitors they send to a site. In Classy's own ChatGPT measurement (40 anonymous runs of 20 business questions from a Turkish IP, 24 Sept 2026), most cited sources were written by someone selling the thing discussed: 63 of 167 citations were platform documentation about the platform's own products and 67 were agency or software-vendor pages; 26 were independent. The sources cited for the same question changed between runs (126 of 143 cited pages appeared only once), while the decision points in the answers stayed much more stable.

How much any single change affects whether ChatGPT recommends a business has not been measured, so we do not promise a ranking or a recommendation.

+Crawl access: not blocking AI crawlers like OAI-SearchBot and GPTBot
+Clear service, pricing and coverage information, in plain, machine-readable language
+Credible third-party presence: directories, press, sector sources
+Real evidence: measured data and case studies, not claims
+Consistent entity information: the same name, address and phone everywhere

We have not measured how much any single item on this list changes whether ChatGPT recommends a business.

For the technical details behind these, see OpenAI's, Google's and Bing's own documentation; we don't rewrite it here, only explain what it means commercially.

What you can't control

By OpenAI's own account, there is no purchasable or guaranteed ranking system for appearing in ChatGPT. So we never say:

×“We'll get you ranked first in ChatGPT”
×“ChatGPT will recommend you”
×“Guaranteed AI visibility”
×“GEO guarantee” or “ChatGPT SEO package”

How does a referral connect to a lead?

There's no single certain way to know a visitor truly came from an AI system, so we use both: observed data (referrer/UTM) and self-reported data (“How did you find us?” on the form). Neither is treated as the sole source of truth.

How do we choose a provider for AI and search visibility, and do we need one?

Choose on evidence rather than promises. No provider can guarantee that ChatGPT or any AI assistant will recommend a business, so an offer that does is the first thing to rule out. Look for a measured baseline before any change, a disclosed method (which questions, how many runs, which dates), results kept even when they are zero, and measurement of enquiries, not only of mentions. Our own measurements are public in the benchmark series below: the first baselines were 0 of 18 and 0 of 19, and later rows record when and where citations appeared. What we do: entity clarity, structured content, machine-readable relationships, source consistency, citation-worthy public information and AI-readable site architecture. Entry is the fixed-price Check-up, which includes an AI-visibility baseline. ClassyDesign sells this service; the benchmark is our own measurement of our own site.

Sources

Benchmark series →

We do not guarantee recommendations, and visibility is not the same as customers: the traffic and enquiries it produces have to be measured separately.

Start with the Check-up.

ClassyDesign AI Discovery Benchmark

This section stays alive, updated here as the numbers change.

Latest · 27 September 2026 · ChatGPT

12/29: Cited in un-branded questions (15 prompts × 2)
4/11: Same question, cited in independent repeats
5/5: Correct firm when asked by name
0/5: Recommended for work we don't do

One day of measurement. We appear for international-client, Gulf, boutique and luxury-brand questions; not in generic “best agency” lists, and rarely for Turkish prompts (1 of 17 valid answers, via our English page). Perplexity still confuses our name with other businesses. Details and raw data in the series below. ↓

First measurement · 18 September 2026

0/18: Mentioned
0/18: Cited
18: Runs measured
2026-09-18: Measurement date

Method: chatgpt.com, no login, web search enabled, 6 distinct customer-acquisition prompts × 3 repeats = 18 runs, Turkish locale, Istanbul context. Each run in a fresh chat; citations read from the real “Sources” panel.

18 Sept 2026 measurement

•For general customer-acquisition and “customers from ChatGPT” questions, ChatGPT cites only official platform sources (Google, OpenAI, Bing); no third-party or agency source appeared.
•On 18 September 2026 ClassyDesign appeared in none of these sources. We did not hide that zero, because it was a real starting point that made later progress measurable; later measurements are in the series below.

Limitations: 18 runs, browser-based anonymous measurement (not API), Turkish/Istanbul context only, a single-date snapshot. The method is repeatable and will be re-run periodically.

A separate layer: in the same measurement window (August 22–September 18, 2026), classydesign.studio's own Google Analytics data measured 4 sessions and 3 first-touch users referred from chatgpt.com. That measures something different from the mention/citation test above: AI Referral Traffic (see section 01), not AI Visibility.

MEASUREMENT SERIES

Every measurement is a new row; old rows are never removed.

The 18 September figures above are the first row. Later measurements are added below with their own date, method and raw data, so what changed and what did not both stay visible.

2026-09-18
Baseline
Methodchatgpt.com, logged out, web search on; 6 Turkish customer-acquisition questions × 3 runs
ChatGPT
Mentioned 0/18 · cited 0/18
Perplexity
Not measured
Starting point.
Limits18 runs, one day; Turkish and Istanbul context only; browser interface, not an API.
Raw data: CSV ↓
2026-09-20
Baseline
MethodRecommendation test: 29 prompts (19 good-fit buyer, 5 wrong-fit control, 5 by name); TR/EN/AR; one run per engine
ChatGPT
Discovery 0/19 · by name 5/5 correct firm · wrong-fit 0/5
Perplexity
Discovery 0/19 · by name 0/5 (confused with other businesses) · wrong-fit 0/5
Invisible in un-branded questions; asked by name, ChatGPT knows us and Perplexity does not.
LimitsOne run per prompt per engine; the engines' model versions are not exposed by the tool.
Raw data: CSV ↓
2026-09-27
Measurement · directly comparable with: 2026-09-20
MethodSame 29 prompts, same tools
ChatGPT
Discovery 1/19 · by name 5/5 correct firm · wrong-fit 0/5
Perplexity
Discovery 0/19 · by name 0/5 (confusion persists) · wrong-fit 0/5
First un-branded match: for “Which Turkish web studios work with international businesses?” ChatGPT listed eight studios; Classy second, source /en/web-design-istanbul. One run; it may not repeat and it is not a recommendation rate.
LimitsOne run; same prompts and tools as 20 September.
Raw data: CSV ↓
2026-09-27 · repeat
Method changed · partially comparable with: 2026-09-27
MethodCoincidence? The same C2 question on ChatGPT as 11 valid independent runs (1 run hit a log-in wall and is excluded); the full 29-prompt set a second time
ChatGPT
C2 cited in 4 of 11 runs (position 2, 2, 3, 6) · full set discovery 0/19 · by name 5/5
Perplexity
Not repeated
Not a one-off: Classy is in ChatGPT's rotating candidate pool for this question and appears in roughly one answer in three; it is not a stable first pick. For comparison, the most frequent studio in the same runs appeared in 8 of 9. Source every time: /en/web-design-istanbul. Asked by name, /en/ai-discovery was cited for the first time.
LimitsOne day; one question; 1 run hit a log-in wall and is excluded.
Raw data: CSV ↓
2026-09-27 · other wordings
Method changed · not directly comparable with: 2026-09-27
MethodDoes it appear for other questions? 15 new un-branded prompts (paraphrases of C2, neighbouring buyer intents, broad category), each twice on ChatGPT; 29 of 30 answers valid
ChatGPT
Cited in 12 of 29 answers · C2 paraphrases 5/10 · neighbouring intents 5/11 · broad category 2/8 · EN 9/19, AR 2/4, TR 1/6
Perplexity
Not measured
Not a quirk of one wording. We appear for international-client, Gulf-company, bespoke boutique and luxury-brand questions, in some of them first in both runs. We do not appear in generic “best agencies” lists, for technical SEO/hreflang intent, or in Turkish. Sources are the English Istanbul web design and luxury brand pages; no TR/AR page was cited. Two answers repeated a “from €3,000” figure removed from the site on 25 September: ChatGPT is reading an older copy of the page.
LimitsEach prompt twice, one day; rates are indicative, not precise.
Raw data: CSV ↓
2026-09-27 · correction
Correction
MethodEvery ChatGPT answer from the same day recounted: the full prompt set twice, 10 extra runs of C2, and 15 variant prompts ×2. 98 answers in total; 75 valid un-branded
ScopeOnly two statements in the notes of the 27 September “other wordings” row. That row's results (12 of 29, TR 1/6) were correct and are unchanged.
CorrectionThe earlier note understated the stale-price observation (“two answers”) and said we did not appear in Turkish. The raw data supports neither; even that row's own result shows TR 1/6. Its sentence “ChatGPT is reading an older copy of the page” also asserted a mechanism we had not shown. What was observed: ChatGPT repeated information no longer present on the live page. · Corrects record: 2026-09-27 · other wordings
Old price statement
In 6 of the 16 un-branded answers that cited Classy: 5 say projects “start around” the starting price removed on 25 September, 1 says we “openly publish” starting prices. All six cited the English Istanbul web design page
Turkish un-branded questions
Cited in 1 of 17 valid answers (not zero); that answer cited Classy's English page
All valid un-branded answers
Cited in 16 of 75 · EN 13/48 · TR 1/17 · AR 2/10
On 27 September the two cited English pages (Istanbul web design, luxury brand) were resubmitted to Bing (IndexNow) for recrawl. The corrected row above stays unchanged.
Limits98 answers, one day. Repeats were run the same day, and source selection can change from run to run. We do not know which cache or index the old information came from, or when ChatGPT will refresh it. The live pages state no project price: projects are priced by scope, and the only fixed-price product is the Check-up.
Raw data: CSV ↓

These are single-run measurements; a difference between two rows is not a trend. No row measures visits, contacts or customers.

In a smaller Classy test (6 runs per topic, 24 Sept 2026, Turkish prompts), when ChatGPT was asked which agency to work with, 14 of 18 cited sources were Turkish agencies' own websites; when asked how to appear in ChatGPT, all 22 cited sources were platform documentation.

This is a small test: six runs per topic on one day; it shows no rate, trend or cause.

Find Customers →

Let's measure what AI assistants cite and send to you

Let's assess together how readable your website is to AI search systems, and where to start.

Assess the Opportunity

What a business can do (effect on AI recommendations not measured)

How much any single change affects whether ChatGPT recommends a business has not been measured, so we do not promise a ranking or a recommendation.

+Crawl access: not blocking AI crawlers like OAI-SearchBot and GPTBot

+Clear service, pricing and coverage information, in plain, machine-readable language

+Credible third-party presence: directories, press, sector sources

+Real evidence: measured data and case studies, not claims

+Consistent entity information: the same name, address and phone everywhere