> ## Content Index > Fetch the complete content index at: https://www.datadab.com/blog/llms.txt > Use this file to discover other available public pages before exploring further. # How to Track AI-Assisted Pipeline in GA4 and Your CRM - URL: https://www.datadab.com/blog/track-ai-assisted-pipeline-ga4-crm/ - Published: 2026-08-31T04:30:00.000Z - Updated: 2026-10-14T07:35:29.000Z - Description: Keep observed AI referrals, buyer-reported discovery, and CRM outcomes separate. A practical GA4 setup with an editable tracking workbook. - Author: Amit Ashwini - Tags: #backlog-2026, #editorial-review *Your attribution report needs a place for the buyer who copied your name out of ChatGPT and arrived through the front door.* The buyer types your domain into their browser. Your report records a direct visit. On the sales call, they say an AI assistant helped them choose whom to contact. One visit. Two useful accounts of it. If you overwrite the first with the second, you lose the distinction between what your systems observed and what the buyer remembers. If you ignore the second, you lose evidence about how the shortlist formed. A neat dashboard can be wrong in either direction. I want both records. I also want the uncertainty left attached. For a B2B SaaS team, that means building a small reporting system around observable referrals, enquiry records, and the buyer's own account. You can do this with GA4 and the CRM you already own. Buying another attribution tool before deciding what those fields mean is an expensive way to postpone a meeting. The [AI-assisted pipeline tracking workbook](https://www.datadab.com/assets/resources/datadab-ai-assisted-pipeline-tracker-v1.xlsx) contains the field definitions, starter referral rules, monthly reporting sheet, and checks described below. It is a setup aid. It contains no measured customer results and does not connect to your accounts. [Download the pipeline tracking workbook](https://www.datadab.com/assets/resources/datadab-ai-assisted-pipeline-tracker-v1.xlsx) ![Editorial cover: AI-assisted pipeline reporting in GA4 and a CRM, with separately recorded evidence meeting at an enquiry.](https://www.datadab.com/blog/content/images/2026/10/track-ai-assisted-pipeline-ga4-crm-01-hero.png) ## Keep observed traffic and buyer testimony apart Start with two different questions. Where did this recorded session come from? What helped this person decide to contact us? GA4 can supply evidence for the first. A prospect can supply evidence for the second. Neither account automatically describes the whole buying process. Google's [documentation on campaigns and traffic sources](https://support.google.com/analytics/answer/11242841) distinguishes user-, session-, and event-scoped traffic-source reporting. Pick the scope that fits the question. Session source is the sensible starting point for a session-referral view; first-user source answers an acquisition question about the user, which may describe an earlier visit. Those two columns can disagree without either being broken. Imagine a prospect first discovered you through a newsletter, returned a week later through a Perplexity link, then booked after a direct visit. That is an illustrative path, not a tracked customer example. A report about the first recorded acquisition and a report about the session that produced the enquiry can each describe a different part of it. The CRM should preserve the prospect's answer alongside those observations. Use a field called something plain, such as “Buyer-reported discovery source”. Store the wording as well as the category. “ChatGPT” and “I checked the comparison in ChatGPT after a colleague recommended you” should not become identical stories merely because both fit the same dropdown. Keep the reporting scope attached. An observed session source and a buyer’s account of discovery are different evidence; overlapping categories must not be added into unique leads. ![A single enquiry record keeps observed session source and buyer-reported discovery as distinct, unfilled evidence fields.](https://www.datadab.com/blog/content/images/2026/10/track-ai-assisted-pipeline-ga4-crm-02-two-evidence-fields.png) ## Build a separate channel view Create an AI-referral view without rewriting the definitions your team already uses for organic search, referrals, or direct traffic. Google supports [custom channel groups](https://support.google.com/analytics/answer/13051316). Its documentation says traffic enters the first channel whose definition it matches in the group's order. Put your specific AI-referral rule before a broader referral rule, or the broader rule can catch the session first. Begin with source values you have actually observed in your property. The workbook includes recognisable domains as starter candidates, including chatgpt.com and perplexity.ai. They are suggestions to validate against your export, not a declaration that every visit from those products will arrive with those source names. Keep the original exported values. Lower-case a copy if your matching process needs it. Do not discard the raw source because a cleaned column looks nicer. An anchored rule that matches a known domain is easier to defend than a broad text search for “AI”. Broad matching can catch unrelated names or miss a changed referral host. Review the unmatched referral rows; that is where new evidence will turn up. For the first reporting pass, I would create the custom group as a separate view, save the original report, and check that a handful of known rows land where expected. Inspect an AI referral, an ordinary partner referral, a direct session, and an organic session. Keep the report dates fixed. There is no reward for deleting the old view on day one. Let the new definition earn your trust. ## Leave Google AI traffic inside its reporting boundary Google's [guidance on AI features](https://developers.google.com/search/docs/appearance/ai-features) says visits from AI Overviews and AI Mode are included in overall Search Console search reporting. You do not get a clean, dedicated breakout simply by naming a spreadsheet column after the feature. Keep that limitation visible. A rise in Google organic visits cannot, on its own, establish a rise in visits from AI Overviews. A screenshot of your page cited in one answer is evidence of that answer, at that time. It cannot supply the missing traffic attribution. Likewise, a direct visit gives you no licence to invent an AI referrer. Direct traffic is a category to inspect alongside other evidence, rather than a convenient bucket from which to extract the number you wanted. The monthly sheet therefore separates observed AI referrals from self-reported AI discovery. Keep all qualified enquiries in your CRM as the wider business total. The categories can overlap; they must not be added into a count of unique prospects. You may have fewer observed referrals than enthusiastic LinkedIn commentary led you to expect. Report the smaller number. The job is to help the business decide what to do, not to provide emotional support for a new marketing channel. ## Give the CRM fields definitions Add a small set of fields your team will use consistently. Existing fields are fine if they already mean the right thing. | Field | What belongs there | What does not | | ------------------------ | ----------------------------------------------------------- | --------------------------------------------------- | | Observed session source | The captured source and its reporting scope | A salesperson's guess | | Buyer-reported discovery | The buyer's answer, plus a useful category | An inferred answer based on the referral | | AI role in the decision | Discovery, comparison, verification, or another stated role | A default assumption that AI created the lead | | Relevant landing page | The page observed or explicitly named | The page marketing hopes influenced them | | Qualification outcome | The agreed sales qualification result | Every form submission | | Evidence note | A permitted quote or report reference | Private conversation data copied without permission | Leave room for “Unknown”. Missing observations and a confirmed zero are different states. For the enquiry form, a short optional question can work: “What helped you decide to contact us?” Let the prospect name an assistant, a peer, a report, or something else. An AI-only question risks turning every answer into a referendum on the thing marketing wants to measure. Sales can ask a follow-up when it is useful: what did the assistant help them work out? Record the answer in ordinary language. You need evidence about the decision, not a questionnaire that makes the first call feel like a border inspection. Keep consent and retention rules appropriate to the data. Do not send names, emails, or private customer statements into public reports just because a worksheet has a spare column. ![CRM schema separates session source, buyer testimony, AI decision role and sales qualification; each retains a definition and unknown values.](https://www.datadab.com/blog/content/images/2026/10/track-ai-assisted-pipeline-ga4-crm-03-crm-field-register.png) ## Treat campaign tags as labels you control Use campaign parameters on links you actually distribute. Google explains how to [collect campaign data with custom URLs](https://support.google.com/analytics/answer/10917952), including source, medium, and campaign parameters. Its guidance also says parameter values are case-sensitive. Write down a naming convention before five people improvise one. A newsletter link to your migration guide might use newsletter as its source, email as its medium, and a named editorial issue as its campaign. A partner's referral link should describe the partner distribution. These tags tell you about a link you supplied. They cannot tell you that an AI assistant independently recommended the page. Do not add campaign tags to every internal link to manufacture more detailed attribution. Keep external distribution labels separate from navigation inside your own site. If a reader clicks from a research article to a diagnostic, the useful observation is the path through those pages; the internal link should not pretend to be a new acquisition campaign. When someone copies a tagged URL into another setting, the original label may travel with it. Record that limitation too. Campaign naming is a helpful discipline, not a witness under oath. The workbook's Checks tab asks whether the team can explain each rule, reproduce the report, and identify where the observation stops. If an answer depends on one person's memory of a meeting, the setup needs another pass. ## Follow the enquiry rather than multiplying it Consider a fictional month with twelve enquiries mentioning an AI assistant. Suppose five of those enquiries also have an observed AI referral. Eight of the twelve meet your qualification rule. You have twelve self-report observations, including five overlapping referral observations, and eight qualified enquiries within the self-report group. You do not have seventeen unique AI-generated leads. The worksheet deliberately avoids a formula that would reward that arithmetic. Record the overlap through enquiry IDs in your private CRM or an approved working file. Keep those IDs out of public case studies. A monthly aggregate can then show the categories without exposing a buyer's details. Add commercial outcomes only when the underlying records support them. Qualified enquiry, accepted opportunity, won customer, and recognised revenue describe different things. Agree the stage definitions with sales and finance. A prospect who says they used ChatGPT and later becomes a customer gives you a useful observation about that customer. It does not prove that ChatGPT caused the sale or that the same result will repeat next month. The difference matters when you decide whether to fund a research article, repair a product page, or buy a monitoring service. A source report, a buyer account, and the actual opportunity can help you make that decision together. They still need their own labels. Our [guide to marketing attribution in B2B](https://www.datadab.com/blog/who-gets-the-credit-how-to-actually-set-up-marketing-attribution-in-b2b/) supplies broader context. For this narrower job, I would favour a small system your team can explain over a sprawling attribution model they cannot inspect. ![Explicitly fictional example: twelve self-reported AI enquiries include five with observed AI referrals; eight of the twelve qualify. The chart does not identify which eight or count referrals outside the group.](https://www.datadab.com/blog/content/images/2026/10/track-ai-assisted-pipeline-ga4-crm-04-count-the-enquiry-once.png) ## Make the report lead to a page decision The strongest observation may be a buyer arriving with the wrong idea about your product. If several buyers say an assistant described your entry plan as including a feature it does not include, inspect your public pricing and comparison pages. If the assistant sent them to an obsolete guide, inspect that guide and its current destination. Keep the original answer or buyer account where permission allows. Choose a page, an owner, and a correction. Save what you changed. Recheck using the same question and conditions where possible, keeping model changes in the record. An attribution spreadsheet should make this sort of work easier. If its only output is a larger slide for the monthly meeting, it has become another content project with excellent stationery. ## Wrap-up I would review the system after one full reporting period. Which sources could we observe? Which buyer accounts added useful context? Could another person reproduce the total? Did any finding change what the team built? Fix the weak definitions before expanding the dashboard. Reliable small counts give finance and sales something they can use. A large blended number gives them something to argue about. Keep the raw observation, the buyer's testimony, and the eventual opportunity attached to the right records. You can learn from all of them without claiming that one field explains the whole sale. Want to get ahead? Use the tracking workbook to find the pages involved in real enquiries, then [check where those pages make the buyer's decision harder](https://www.datadab.com/tools/decision-friction-scanner). [Scan your page for decision friction](https://www.datadab.com/tools/decision-friction-scanner)