Isolating AI Assistant Referrals

Assistant referrals arrive with inconsistent referrer values, some carry no referrer at all, and default channel groupings drop them into direct, referral or other. A saved segment matching two or three known hostnames goes stale quietly, and it counts none of the visits where somebody read an answer and then typed your brand into a browser. Measure what can be attributed, size what cannot, and label the difference rather than reporting one number as though it were complete.

The skill file

---
name: ai-referral-session-isolation
description: Use when someone asks how much traffic comes from AI assistants and the channel grouping has scattered it across direct and referral.
---

# Isolating AI Assistant Referrals

Assistant referrals arrive with inconsistent referrer values, some carry no referrer at all, and default channel groupings drop them into direct, referral or other. A saved segment matching two or three known hostnames goes stale quietly, and it counts none of the visits where somebody read an answer and then typed your brand into a browser. Measure what can be attributed, size what cannot, and label the difference rather than reporting one number as though it were complete.

## What you need first

- referrer-level session data, not the channel grouping report
- server access logs for the same period, which retain referrers analytics loses
- a maintained list of assistant referrer hostnames and their URL parameters
- branded search and direct traffic baselines from before the period in question

## Method

1. Pull raw referrer values rather than the channel report, because assistants land in whatever bucket the default grouping guessed. Read the long tail by hand once so you learn what these referrers actually look like on your site.
2. Build a match rule from hostnames and, where present, query parameters, and keep it in the same versioned file as your other segment definitions so it is reviewed rather than forgotten.
3. Create a new channel or segment instead of editing an existing one, so historic reports do not change retroactively and last quarter still reconciles.
4. Cross-check one week against server logs. Analytics loses referrers through redirect chains and privacy settings, and the log is what tells you how large that loss is.
5. Separate assistant referrals from assistant crawlers in the log analysis. They often share infrastructure, and mixing them turns a bot fetch into a reported visit.
6. Size the invisible portion rather than ignoring it: track branded search and direct traffic to the pages assistants cite, and report any movement there as indicative and clearly labelled as such.
7. Report volume next to behaviour, and say plainly when the base is too small. At most sites these session counts are still low enough that a conversion rate on them is not yet a reliable number.

## What this produces

An AI referral segment with a versioned match rule, a log-based coverage estimate, and a short statement of what the segment cannot attribute.

## Where this goes wrong

- Reporting a percentage change on a base of a few dozen sessions, which reverses next month and costs credibility for the rest of the report
- Letting the referrer list go stale, so a flat line reads as a plateau when it is really a rule that stopped matching
- Counting crawler fetches as referral sessions, which on some sites inflates the figure by an order of magnitude
- Attributing every rise in direct traffic to AI, when direct also absorbs app clicks, untagged email and dark social

---

From the QuQi skill library - https://www.quqi.io/skills/ai-referral-session-isolation

What you need first

• referrer-level session data, not the channel grouping report
• server access logs for the same period, which retain referrers analytics loses
• a maintained list of assistant referrer hostnames and their URL parameters
• branded search and direct traffic baselines from before the period in question

Method

1. 01 Pull raw referrer values rather than the channel report, because assistants land in whatever bucket the default grouping guessed. Read the long tail by hand once so you learn what these referrers actually look like on your site.
2. 02 Build a match rule from hostnames and, where present, query parameters, and keep it in the same versioned file as your other segment definitions so it is reviewed rather than forgotten.
3. 03 Create a new channel or segment instead of editing an existing one, so historic reports do not change retroactively and last quarter still reconciles.
4. 04 Cross-check one week against server logs. Analytics loses referrers through redirect chains and privacy settings, and the log is what tells you how large that loss is.
5. 05 Separate assistant referrals from assistant crawlers in the log analysis. They often share infrastructure, and mixing them turns a bot fetch into a reported visit.
6. 06 Size the invisible portion rather than ignoring it: track branded search and direct traffic to the pages assistants cite, and report any movement there as indicative and clearly labelled as such.
7. 07 Report volume next to behaviour, and say plainly when the base is too small. At most sites these session counts are still low enough that a conversion rate on them is not yet a reliable number.

What this produces

An AI referral segment with a versioned match rule, a log-based coverage estimate, and a short statement of what the segment cannot attribute.

Where this goes wrong

• Reporting a percentage change on a base of a few dozen sessions, which reverses next month and costs credibility for the rest of the report
• Letting the referrer list go stale, so a flat line reads as a plateau when it is really a rule that stopped matching
• Counting crawler fetches as referral sessions, which on some sites inflates the figure by an order of magnitude
• Attributing every rise in direct traffic to AI, when direct also absorbs app clicks, untagged email and dark social

Use this skill in your own AI

The download is a plain markdown file with the name and trigger in its frontmatter. Where an assistant supports skills it can load itself, that frontmatter is what it reads to decide this one applies.

Questions about this skill

More in Analytics & reporting