# The buyer says ChatGPT, Pipedrive says Web forms: keep both and read the gap

https://salesbond.app/blog/self-reported-attribution-pipedrive/
Published: 2026-10-07 · Updated: 2026-10-07

Keep what the system recorded and what the buyer said in two separate fields. This is self-reported attribution: asking buyers how they found you and storing the answer the way they said it. Pipedrive fills source origin on its own and nobody can edit it; the buyer’s words, “I asked ChatGPT” or “found you on Perplexity”, go into a plain text field that reps do not rewrite. Once a month you put the two columns next to each other and look at where they disagree. The disagreement is the report. This article is about the second field and the monthly read; the first field and the source list are covered in [how web forms fill the Pipedrive lead source](https://salesbond.app/blog/pipedrive-lead-source-web-forms/), and we won’t repeat that here.

## A lead says “ChatGPT” and your CRM says “Web forms”

The team below is hypothetical. The call is the pattern this article is about.

Fintan is four minutes into a first call. The buyer, Seraphine, runs sales at a twelve-person agency, and before he can ask his second question she answers one he did not ask: “I was asking ChatGPT which tools check Pipedrive for stalled deals and it listed you, so I filled in your form.” He says great, writes nothing down, and moves on to budget.

The deal he is looking at while they talk already has a source. The sidebar says source origin: Web forms, source channel: Inbound website. Both are correct. She did come through the form on the website. Pipedrive saw the form and recorded the form, which is exactly what it is built to do.

At the end of the month Wilhelmina, who runs the team, counts deals by source and sees that the website is the best channel the company has. It is, in the sense that the website is where the deals arrive, including every deal that started in a chat window, because the assistant hands people a link and the link goes to the form. Nothing in Pipedrive is wrong. The one sentence that would have changed the budget conversation evaporated on Fintan’s call.

The CRM knows the pipe. Only the buyer knows the reason, and the buyer told you, once, out loud.

## What Pipedrive records on its own, and what it can’t know

Pipedrive gives every deal and lead a set of source fields, and only half of them are yours to edit. Source origin is how the record entered Pipedrive (Web forms, Import, API and so on); Pipedrive fills it and nobody can change it in the interface. Source channel is a list of values an admin maintains ( [Pipedrive: lead source in deals](https://support.pipedrive.com/en/article/lead-source-deals), updated September 3, 2026). Our [lead source article](https://salesbond.app/blog/pipedrive-lead-source-web-forms/) walks through all four fields and how to set them up.

Look at how the help page defines source channel: “the sales or marketing tool that added data to Pipedrive.” A tool. Every value Pipedrive lists is a mechanism. There is no row for “a person asked a language model and it recommended us”, and there shouldn’t be, because that is not how the data got into Pipedrive. The data got in through the form.

So the tempting fix, adding “ChatGPT” as a source channel option, puts an answer in a field built for a different question. It also mixes what the system observed with what a human remembered in one column. Once they share a dropdown, nobody reading the report can tell which one they are looking at. Keep the channel list what that article makes it: short, and about the channels you spend money on. The buyer’s story needs its own field.

## One extra field: the buyer’s own words

The field is a text field on the deal, and it holds exactly what the buyer said, unedited. No category, no rep’s interpretation. If Seraphine said “asked ChatGPT for CRM add-ons, it listed you”, that is the value.

Here is the setup, with the plan notes from Pipedrive’s own help so you know what your account will show:

1. Open the account menu in the upper right corner, go to Personal preferences > Data fields, pick the Lead/Deal entity and click “+ Custom field”. Pipedrive creates the same field for leads and deals in one step ( [Pipedrive: custom fields](https://support.pipedrive.com/en/article/custom-fields), updated September 3, 2026).
2. Name it so the purpose is in the name: “How did you hear about us (buyer’s words)”. Type: Text. One line is enough; most answers are a short phrase.
3. Write the field description, available on the Growth plan and above, and make it one instruction: “Paste what the buyer actually said. Do not clean it up or shorten it.” Reps see it behind the info icon next to the field ( [Pipedrive: custom fields](https://support.pipedrive.com/en/article/custom-fields) ).
4. On Premium, use the read-only setting to decide which users can edit the field. Lock it because a tidy rep will “fix” the text into a category, and a category is the one thing this field must not become. Read-only still leaves doors open to automations and imports; [the custom field cleanup article](https://salesbond.app/blog/pipedrive-custom-fields-cleanup-data-fields/) lists them.
5. Add the field to your web form as an input block. Web forms accept Text and Large text custom fields. The “Required” checkbox only appears for single-option custom fields, or multiple-option custom fields with only one available input, so a text field stays optional on the form in any case ( [Pipedrive: Web Forms](https://support.pipedrive.com/en/article/web-forms), updated September 3, 2026).

Optional is fine. Its what we would pick anyway. The one audit we have seen with real numbers made the field a single-line text, last on the form, not required, and 62% of people answered it. In that audit of 8,307 HubSpot contacts, the CRM’s own original source field attributed 11 leads to AI tools, while the free-text “How did you hear about us?” answers named ChatGPT, Claude, Perplexity, Gemini or just “AI” for 809 buyers ( [Rock the Rankings: SaaS AI search attribution](https://www.rocktherankings.com/playbooks/saas-ai-search-attribution)). That is one company’s data, published by an SEO agency, so do not treat it as a market figure. Treat it as proof that a free text box can see something the system fields structurally cannot.

## Normalizing the answers once a month

Free text does not count, so once a month somebody turns it into categories, in a second field, and never by editing the first. The attribution guides we read agree on the shape: keep the measured source and the stated source in separate fields, normalize the free text into buckets, reconcile monthly ( [Lead Journey: self-reported attribution](https://leadjourney.io/glossary/self-reported-attribution); [Prose Media: marketing attribution in 2026](https://www.prosemedia.com/blog/marketing-attribution-2026-ga4-crm)). The second field is a single-option deal field, call it “Source (buyer’s words, normalized)”, and our rules for it are short:

- Keep the option list under ten. “AI assistant” is one bucket. Splitting it into ChatGPT, Claude, Perplexity and Gemini gives you four thin columns and nothing you can act on this year.
- One person normalizes, in one sitting, with the deal list filtered to deals created last month where the raw field is not empty. For a small team that is a short sitting, not a project.
- When the raw text names two things (“a friend mentioned you and I checked on ChatGPT”), pick the one the buyer said first. Write that rule down so the next person applies it the same way.
- The raw field is never touched. If you cannot tell what the buyer meant, the normalized value is “Unclear”, not a guess.

Two fields, one list, one monthly sitting. An attribution routine a team of six will actually do beats a sophisticated one they abandon in March.

## The gap is the finding: reading tracked vs. said

Where the two fields disagree is what your dropdown can’t see. Open the deal list for last month, add two columns, source origin and your normalized field, and sort by the normalized one. Which column is right doesn’t matter here. The rows where they disagree do.

![Side-by-side columns of system source and buyer-reported source for eight deals, with the mismatched rows highlighted](https://salesbond.app/wp-content/uploads/2026/10/self-reported-attribution-pipedrive-2.png)

In the HubSpot audit above the gap was 11 against 809 for AI tools ( [Rock the Rankings](https://www.rocktherankings.com/playbooks/saas-ai-search-attribution)). Yours may be smaller, and the shape may differ. The reading is the same either way. Rows where the system says Web forms and the buyer says a referral tell you a customer is sending people to your site, and the [referral program](https://salesbond.app/blog/referral-partner-deals-pipedrive/) deserves a line in the budget. Rows where the system says Web forms and the buyer says an AI assistant tell you the assistant is the one doing the recommending, and someone should go find out what it says about you. Rows where the two agree are teh easy part.

Do not use the gap to “correct” anything. Source origin cannot be edited, and source channel should not be rewritten to match the buyer, for the reason above. Two fields that each tell the truth about a different thing are worth more than one field that averages them.

A disclosure, since this is the place where tools get mentioned. We build SalesBond, a sales management layer on top of the CRM a team already uses; we started with Pipedrive. A field like this one is only useful if it is actually filled, and in a small team the hard part of [CRM data quality](https://salesbond.app/blog/crm-data-quality-metrics/) comes in week two, when new deals start arriving with it blank. If you want a second pair of eyes on fields like that, [check your CRM with SalesBond](https://salesbond.app/). The two-column read itself needs nothing but a filter.

## If the form never asked: one question on the first call

Plenty of your deals will never see the form. They arrive by referral, by email, from an event, from a rep’s outbound list. For those, the rep fills the field on the first call, in the buyer’s words.

The question is one sentence and it goes early in the [discovery call](https://salesbond.app/blog/discovery-call-small-team-pipedrive/), before the agenda, when people are still telling you how they got there anyway: “Before we start, how did you first come across us?” If the answer is a mechanism (“your website”), one follow-up: “Was it a person or a tool that pointed you there?” That second question is where “oh, I asked ChatGPT” lives. Buyers do not volunteer it because it does not feel like an answer to “how did you hear about us”. It is the answer.

Two things make this hold. The rep types while the buyer talks, into the raw field, verbatim; keeping the deal open in a second window during the call makes that easy. A note written after the call becomes “AI” or “website” because the rep is summarizing, and summaries are what the raw field exists to avoid. And the question is asked every time, including when the the form already answered it. A buyer who typed “ChatGPT” into the form and says “a colleague” on the call just gave you a second data point.

Should it be a required field? No. Marking it required at a stage would block the deal until somebody types something, and what they type under pressure is “n/a”. [Required fields in Pipedrive](https://salesbond.app/blog/required-fields-pipedrive/) have their place, this field is not one of them. The fill rate is a management question, and a rep who leaves it blank for a month gets asked about it in a one-to-one, not blocked by the software.

## What self-reported attribution can’t tell you

Self-reported attribution has two known limits and this setup does not remove either of them.

The first is recency. People remember the last memorable thing before they acted, not the first time they heard of you ( [Lead Journey: self-reported attribution](https://leadjourney.io/glossary/self-reported-attribution)). Seraphine asked ChatGPT last week. She may have read a post of yours in March and forgotten it. “AI assistant” in the normalized field is honest about what she remembers, and silent about what she does not.

The second is that both fields are single-touch. Source origin is the last mechanical touch: the form. The buyer’s words are the last memorable touch: the assistant. A deal that went article, newsletter, colleague, ChatGPT, form gets recorded as two of those five. The two-column read shows where your mechanism-only view is blind. It does not give you the path and a team of six does not need the path. It needs to know whether the thing it keeps hearing on calls is real at the level of the month.

When Wilhelmina reads the two columns at the end of the month, she still won’t know what the assistant said, or why, or whether it will say the same thing next month. She will know something she did not know the month before, and it cost her one text field.

So add the text field this week, ask the question on your next first call, and put the two columns side by side when the month closes.