# The only way to know ChatGPT sent them is to ask

_Use case · AI assistants_

Canonical: https://www.humansurvey.co/use-cases/ai-assistants

ChatGPT, Claude, Perplexity and Gemini do not send a referrer. The people they send you arrive as `direct / none`, in the same bucket as someone typing your domain from memory. The figure that gets quoted is around seventy percent of AI-assistant referrals landing there. Treat the exact number as folklore if you like — **the mechanism is not folklore, and no log-based method fixes it.** One row in a candidate list does.

---

## Why it lands in Direct

Four separate things have to go right for a referrer to survive, and with an assistant in the middle they mostly do not:

- The answer is rendered in a native app or a desktop client, which is not a web page and has no referrer to pass on.
- Links are opened in an in-app browser, or carry `rel="noreferrer"`, or route through a redirector that drops the header.
- The citation is a link to somebody else — a listicle, a Reddit thread, your own docs page — so whatever referrer does arrive names that page and not the assistant.
- **Most often, there is no click at all.** The person reads a name in an answer, searches it, and lands on you from Google. That path is not a tracking gap you could close with better instrumentation — the visit genuinely came from search. The _discovery_ happened somewhere with no log.

That last one is the whole argument for asking. A perfect referrer header would still book this person against Google, because Google is where they clicked. Only the person knows the assistant was the reason, and they will tell you if you make it a single tap.

## The wording is most of the work

This channel is more sensitive to phrasing than any other, for a reason worth being precise about: people do not experience themselves as having come from an LLM. They asked a question, got a name, searched the name, and remember Google.

> "AI assistant" is a category. "ChatGPT" is a memory.

- **Use the product name, never the category.** A brand name has a logo, a colour and an app icon behind it, and costs the respondent no translation. A category asks them to classify their own experience first, which is exactly the step that ends in _I don't remember_.
- **Put it beside Google, not under it.** Nesting the assistants inside a "search" group makes the respondent decide whether asking ChatGPT counts as searching. The catalog's default channel list ships `chatgpt` as its own row, next to `google`, for that reason.
- **One row per assistant, not one row for all of them.** Four rows cost four lines of config and are scanned, not read — a list with logos is affordable in a way a list of sentences is not. Collapse them into one and you can never answer "is this ChatGPT or is this all four", which is the first question anyone asks of the number.
- Order rotates per respondent by default, so no assistant sits above Google for everybody and the raw share is unbiased by construction. If you pin the order with `"order": "fixed"`, you inherit the position bias that comes with it.

One production detail rather than a surprise later: ChatGPT's mark is absent from the icon set the product generates from, following a trademark request, so that row renders as a monogram tile instead of a logo. Claude, Perplexity and Gemini have marks. The catalog tells you which is which before you configure anything — `GET /api/attribution/catalog` needs no key.

## Where the question goes

Lead with the **payment or upgrade flow**. The respondent has just paid, so the answer is joined to revenue with no conversion tracking at all, and the confirmation screen was dead space anyway. "ChatGPT produced this much revenue last month" is the sentence a budget holder acts on.

Then run a second form in the **signup flow**. It is the only way to see the people an assistant sends who never pay — and with both running, the same channel's share among payers versus among signups _is_ its signup-to-paid conversion rate. Nothing else reports that, because nothing else asks twice.

Ask early inside each flow. Memory decays, and asking late means asking only the people who stayed: if a channel sends users who churn in week one, a late-placed question systematically under-counts it. A small sample is visibly small. A biased one is not.

## The follow-up: what they were asking

Picking a candidate can expand a follow-up in place, no page transition. Point the assistant rows at a node that asks what they were asking about, and what comes back is **a real question from a real person who then converted** — a different artifact from the prompt sets an AI-visibility tool guesses at and scores you against. Nobody has to guess which prompts matter when the people who bought tell you theirs.

Nothing ships preconfigured here. The assistant catalog entries carry no follow-up by default, and the topic list is yours, because only you know what your buyers ask. This is you spending one extra tap on the channel you care about most this month:

```json
{
  "nodes": [
    {
      "id": "channel",
      "prompt": "Where did you first hear about us?",
      "order": "rotate",
      "allow_free_text": true,
      "candidates": [
        { "id": "google",     "catalog_slug": "google" },
        { "id": "chatgpt",    "catalog_slug": "chatgpt",    "expands": "ai_topic" },
        { "id": "perplexity", "catalog_slug": "perplexity", "expands": "ai_topic" },
        { "id": "claude",     "catalog_slug": "claude",     "expands": "ai_topic" },
        { "id": "gemini",     "catalog_slug": "gemini",     "expands": "ai_topic" },
        { "id": "reddit",     "catalog_slug": "reddit" },
        { "id": "friend",     "catalog_slug": "friend" },
        { "id": "dunno", "label": "I don't remember",
          "pinned": "end", "dont_remember": true }
      ]
    },
    {
      "id": "ai_topic",
      "prompt": "What were you asking about?",
      "allow_free_text": true,
      "candidates": [
        { "id": "topic_alternatives", "label": "Alternatives to a tool I already use" },
        { "id": "topic_howto",        "label": "How to do a specific thing" },
        { "id": "topic_pricing",      "label": "Comparing prices" },
        { "id": "topic_dunno", "label": "I don't remember",
          "pinned": "end", "dont_remember": true }
      ]
    }
  ]
}
```

- `allow_free_text` is where the value actually is. Your three topic buckets are a guess; what someone types is the phrasing. It is stored verbatim, capped at 500 characters, and never truncated silently.
- All four assistants expanding into one `ai_topic` node pools the topic counts across them. Give each its own node if you need topics per assistant — the rollup keys rows on node × candidate, so pooling is a decision you make in the config, not one you can undo at read time.
- A follow-up costs a tap, so it belongs on the channels carrying spend. The respondent who answers the first question and walks away has still told you the channel, and that response counts.

## What comes back

There is no dashboard. Your agent reads the aggregate over HTTP and writes the note:

```bash
curl "https://www.humansurvey.co/api/attribution/rollup\
?form_id=abc123efgh45&by=candidate&metric=revenue&from=2026-07-01&to=2026-08-01" \
  -H "Authorization: Bearer hs_sk_..."
```

```jsonc
// illustrative — one month of a payment-flow placement
"denominator": { "completed_responses": 1204, "per_node": { "channel": 1204, "ai_topic": 118 } },
"rows": [
  { "node_id": "channel", "candidate_id": "google",     "responses": 388, "share": 0.32, "revenue_cents": 1610000 },
  { "node_id": "channel", "candidate_id": "chatgpt",    "responses":  96, "share": 0.08, "revenue_cents":  486000 },
  { "node_id": "channel", "candidate_id": "perplexity", "responses":  21, "share": 0.02, "revenue_cents":   92000 },
  { "node_id": "channel", "candidate_id": "claude",     "responses":  14, "share": 0.01, "revenue_cents":   71000 },
  { "node_id": "channel", "candidate_id": "gemini",     "responses":   9, "share": 0.01, "revenue_cents":   38000 },
  { "node_id": "ai_topic", "candidate_id": "topic_alternatives", "responses": 51, "share": 0.43, "revenue_cents": null }
],
"unresolved": { "raw": 61, "dont_remember": 143, "skipped": 88, "per_node": { … } }
```

Two things about that payload matter more than the row values. The **denominator ships next to the shares**, so 8% is 96 of 1,204 and not a percentage of some population you have to infer. And the people who did not give you an answer stay visible in `unresolved` instead of being dropped from the base — dropping them would inflate every channel on the page, ChatGPT included.

Free text arrives unresolved on purpose. Once a month, read what piled up and map the recurring phrasings onto a topic — retroactively, across every window that already went out:

```bash
curl "https://www.humansurvey.co/api/attribution/forms/abc123efgh45/unresolved?node_id=ai_topic" \
  -H "Authorization: Bearer hs_sk_..."
# → the phrasings people typed, grouped and counted, most frequent first:
#   "best invoicing tool for freelancers"        7
#   "alternative to <competitor>"                5
#   "how to send a quote that gets signed"       4
```

A mapping is not an edit. Nothing about the stored response changes; the rollup resolves against the live remap table on every read, so one row fixes two months of history at once and revoking it moves them back.

## What this does not tell you

The product's job is not to hand you a confident percentage. Most attribution tools do that, and it is why nobody believes them. So, plainly:

- **It is self-report, and self-report is imperfect.** People misremember, and some of them will have met your name twice. What the shipped numbers do is refuse to hide that: the base is always there, and _I don't remember_ is a first-class row rather than a rounding error.
- It measures _where they first heard_, not what closed them. Last touch is near-constant — people search your brand name — and carries no budget decision.
- "ChatGPT" is not proof the assistant recommended you. It could have cited an article about you, and the respondent cannot tell the difference. Read the row as _this person's discovery ran through ChatGPT_, which is still the thing you were unable to see at all.
- `share_corrected`, `position_effect` and `calibration` come back as explicit nulls today, not as computed numbers. Rotation already makes the raw share unbiased, so the correction would only serve someone who pinned the order — and an absent number is better than a smoothed guess.
- Small channels are small samples. Nine Gemini responses is nine responses, and the payload gives you the count so you can decline to draw a conclusion from it.

## Getting started

Sign in, copy a key, hand it to your agent. From there it reads the catalog, creates one form per placement, writes the candidate lists, and gives you a URL to embed early in checkout and in signup. A month later you ask it how the month went.

## More

- Docs — endpoints, embed contract, full rollup shape: https://www.humansurvey.co/docs
- Agent-readable overview: https://www.humansurvey.co/llms.txt
- Full machine reference: https://www.humansurvey.co/llms-full.txt
- Communities and word of mouth: https://www.humansurvey.co/use-cases/community-feedback
- Launch day: https://www.humansurvey.co/use-cases/product-launch
- Podcasts and events: https://www.humansurvey.co/use-cases/events