ChatGPT out-earned our Google Ads by 6.7x

Why almost nobody else can publish this

Most businesses report close to zero AI traffic because of how they measure, not because of what happened. ChatGPT strips the referrer on many links, its mobile apps pass nothing at all, and a large share of people read an answer, close it, and type the company name into a browser. Every one of those lands in a Direct bucket, or nowhere.

So a business measuring referrer alone sees a rounding error and concludes AI sends no customers. Panda Patches captures the referrer where it survives and asks the customer where it does not. Both answers feed one source field. That is the only reason these numbers are not zero, and it is the single most transferable thing in this article.

Revenue by source, 1 March to 18 August 2026

Every row Panda Patches recorded, including the ones that produced nothing. Sorted by revenue. AI assistants are marked so you can find them without hunting.

Both totals exclude Web Checkout (48 orders, $10,295.11), which the Panda Patches operations platform treats as a fulfilment route rather than an acquisition source and already excludes from its own revenue total. Excluding it from the order count too keeps both figures describing the same population; mixing them was an error caught by summing the rows against the declared totals.

What each AI assistant produced

ChatGPT did the volume, but it is not the whole picture. Six assistants appear separately, including Google AI Overview, which most analytics setups cannot distinguish from ordinary Google organic traffic at all.

The average order across the whole store is $357.40. ChatGPT sits below that at $227.74, so it brings more orders of smaller size. Claude sits well above it at $630.00. That Claude figure rests on 4 orders and proves nothing. It is here because leaving it out would be selective, not because four orders support a conclusion. Treat it as a thing to watch, not a finding.

The channels that produced nothing

TikTok, YouTube, LinkedIn each recorded zero orders and zero revenue across the whole window. Those are real zeroes, published as zeroes.

We include them because a table showing only the sources that worked is an advertisement. The useful version shows what was tried and did not pay, which for this store is the entire short-form video category over six months.

Google's own instrument agrees

A second, unrelated measurement points the same way. Google Search Console, Performance, Generative AI features (Beta) for the same property recorded 48,900 impressions over 3 months to 18 August 2026, rising from roughly 400 a day in mid-May to 1121 on 2026-08-12.

That matters more than either number alone. The store's own attribution could be flattering itself, and Google's report knows nothing about the store's orders. Two instruments with no connection to each other showing the same direction is the difference between a claim and evidence. Impressions in Google's generative surfaces, not clicks and not revenue. It corroborates direction, not the dollar figures.

What a customer actually typed

One customer explained where she had come from without being asked. She had gone to ChatGPT and asked for patch companies with no or low minimums, and the store came up.

The query is the useful part. She did not search the category, she searched a constraint. Nobody wins that phrasing by publishing a page about custom patches; you win it by stating your minimum order quantity somewhere a model can read it. Every product decision the business had already made and never written down plainly is a query like that one going to a competitor. We are quoting her anonymously because permission to use her name is hers to give and had not been given when this was published.

How this compares to the one large study that exists

A 94-brand study found the same thing about order value that a single store found here. Visibility Labs analysed twelve months of Google Analytics data from 94 seven and eight-figure ecommerce brands across 2025, comparing 9.46 million non-branded organic sessions against 135,000 ChatGPT referral sessions. ChatGPT traffic converted at 1.81% versus 1.39% for non-branded organic, but its average order value was $204 against $238 for organic, a 14.3% gap, leaving revenue per session 10.3% higher rather than 31%.

Our single store shows the same direction on the part that overlaps. ChatGPT's average order here is $227.74 against a site average of $357.40. Two samples with nothing in common, one of 94 brands and one of 760 orders, both finding that AI assistants send smaller orders than the channel they are compared against. That is worth more than either finding alone, and it is the reason the AOV column is on this page at all.

What is not in that study, or any other we could find, is revenue by individual assistant. Published work reports ChatGPT against organic search. It does not separate Claude from Perplexity from Google AI Overview, because almost nobody can. That separation is the part of this dataset with no equivalent elsewhere, and it is only possible because the source question is asked at the point of enquiry rather than inferred from a referrer header.

What this does not establish

This is one business, in one vertical, over one window. One business, one vertical (custom embroidered and printed patches), one window. These are not industry benchmarks and no claim is made that they generalise. A store selling custom patches to people who need small runs is close to an ideal case for an assistant that answers constraint-shaped questions. Your category may behave nothing like it.

There are no conversion rates on this page, deliberately.The lead records and the order records do not share a window, and repeat orders appear to inherit the original lead's source. Dividing one by the other produced a conversion rate above 100% for two sources, which is impossible, so the whole measure is withheld rather than presented in a softened form.

One outlier is disclosed rather than removed. Average order value of $6,120 against a site average of $348.91, almost certainly bulk or wholesale. Removing it moves the AI revenue share from 6.2% to 6.5%, so it works against the headline rather than for it.

And self-reported attribution has its own bias. A customer who used ChatGPT and then searched Google may report either one. The honest reading is that these figures are better than referrer-only tracking and still not exact, which is true of every attribution number anyone has ever published.

What we would do with this if it were your business

Add the how-did-you-hear question to your enquiry form today. It is the cheapest change on this list and without it none of the rest is visible. Then wait a quarter before drawing any conclusion, because a single large order will swing a small channel's share by several points.

Then write down the constraints you already satisfy. Minimum order quantity, turnaround, materials, what you will not do. Those are the terms the questions are shaped around, and most businesses keep them in a salesperson's head. Our answer engine playbook covers the structural side, and agentic browsing covers what an assistant has to be able to load before any of it counts.

Frequently asked questions

Method

Source: Custom operations platform built on Next.js, Sanity, Supabase and Square. Attribution: Referrer captured where it survives; self-reported source collected at quote request where it does not. Both fields feed one source value. Window: 1 March to 18 August 2026, read on 2026-08-18. Relationship: Founder-affiliated. Owned and operated by a PandaCodeGen co-founder. Corroborating instrument: Google Search Console, Performance, Generative AI features (Beta), 3 months to 18 August 2026.

The underlying figures live in a single data file in this site's repository, and every percentage on this page is computed from it at build time rather than typed into a sentence. If a row and a total ever disagree, the build fails rather than publishing two numbers that contradict each other.