How to Attribute AI Visibility Changes Without Inventing Causation

The tempting story is usually too confident

The team adds a compatibility table on July 1. On July 8, ChatGPT cites the product page.

The tempting conclusion:

The compatibility table caused ChatGPT visibility.

What else changed?

• product returned to stock
• feed refreshed
• competitor went out of stock
• prompt wording differed
• ChatGPT mode or model changed
• search index refreshed
• a third-party review was published

The timing is useful evidence. It is not enough by itself to establish causation.

Build a change graph

The attribution system connects:

Product and variant
-> diagnosed gap
-> approved fields
-> deployment
-> public verification
-> feed and channel events
-> stable prompt runs
-> answer facts and citations
-> referral or order events

Every edge has a timestamp and evidence source.

Use stable comparison units

Keep:

• prompt ID and exact wording
• provider and visible mode
• country and language
• product and market
• completed run status
• source URLs
• expected facts
• named competitors

Compare completed runs only. A timeout cannot become a zero.

Define attribution levels

Level 1: temporal association

The observation happened after the change.

Defensible statement:

The canonical citation was first recorded seven days after the compatibility release.

Level 2: matched evidence

The later answer uses the corrected fact or newly published source.

Defensible statement:

The answer stated the DisplayLink condition added in the July 1 release and cited the updated compatibility page.

Level 3: controlled test-set change

Stable prompts and comparable periods show a consistent difference.

Defensible statement:

Canonical citations increased from 4 of 18 completed runs to 7 of 19 completed runs across the fixed test set. Two inventory changes and one provider update also occurred.

Level 4: commerce attribution

A legitimate referral or order record carries channel attribution.

Defensible statement:

Three orders in the period carried the available AI-channel attribution. This measures attributed transactions, not every influence on the purchase.

Do not hide conflicting evidence

If citations improved but fact accuracy declined, report both. If a provider update occurred, include it. If prompts changed, do not present the periods as equivalent.

Attribution should be able to say:

Inconclusive

That result is better than a fabricated ROI claim.

Connect attribution back to operations

The final record should answer:

• what changed
• why it changed
• who approved it
• whether it deployed correctly
• which observations followed
• which corrected evidence appeared
• what else changed
• what claim the evidence supports

The result returns to Monitor as the new baseline. Attribution is not the end of a campaign. It closes one cycle in an ongoing product visibility operation.