AI Search Attribution Models Compared: What to Use

AI search attribution models help you understand whether ChatGPT, Perplexity, Claude, Gemini and Google AI Overviews contribute to sales and leads. No single model captures the whole journey: use last click for the revenue you can verify, multi-touch for observed journeys, surveys for discovery that analytics misses, and incrementality testing to estimate causal impact. For a broader view of the metrics behind AI visibility, see our guide to AEO KPIs.

The reason is simple. AI answers can influence a buyer without sending them to your website. When a visitor does arrive, the referrer may not identify the AI platform, and analytics can label the session as Direct, Organic Search or Referral. A dashboard that reports only recorded AI referrals will therefore show a useful but incomplete picture.

The practical answer is to combine methods, keep their results separate, and state what each one can and cannot prove. This comparison explains the four main approaches, how to implement them, and which combination fits an agency, B2B SaaS business or e-commerce brand.

What AI search attribution models measure

Attribution is the method a business uses to assign credit for a conversion to marketing interactions. A conversion might be a purchase, a qualified lead, a booked demo or another business outcome.

Traditional website attribution depends on observable interactions: an ad click, a search result click, a referral session or an email visit. AI search complicates that picture. A person can ask an assistant for recommendations, read its answer, and then visit your site by typing your brand name or searching for it separately. The AI influence happened, but the visit may not carry an AI referrer.

There are two different questions to keep apart:

• What conversions can we directly observe coming from AI search? Analytics and CRM records can help answer this.
• How much did AI search change demand or sales overall? Surveys and experiments can help estimate this, but they do not identify every individual journey.

The first is a measurement question. The second is a causal question. Confusing them leads to inflated claims or, just as often, to the false conclusion that AI search had no effect.

For context, Digital Applied’s analysis of the SparkToro study reports that 68% of Google searches end without a click. That figure refers to Google searches, not specifically to AI answers, but it illustrates why click-based analytics cannot represent every search interaction. Read the analysis of the zero-click study.

The four models side by side

These approaches are not interchangeable. Last click and multi-touch allocate credit among recorded interactions. Surveys collect a buyer’s account. Incrementality compares outcomes with a counterfactual: what might have happened without the work being tested.

A useful reporting approach combines the evidence without blending it into a single made-up number. Show recorded conversions separately from survey responses and modeled lift. SearchFIT’s resources can help your team build a measurement plan alongside its AI visibility work.

Last-click attribution: the observable floor

How it works

Last-click attribution gives all conversion credit to the final recorded source or session before a purchase or lead submission. If someone clicks a link from Perplexity, lands on your website and converts in that session, an analytics tool may record a referral or AI-related visit and award it the conversion.

It is popular because it is straightforward. The events are in analytics, the conversion can be checked against an order or CRM record, and the result is easy to explain. For that reason, last click is valuable as a verified floor: the conversions you can connect to a recorded AI-related visit.

It is a floor, not a full estimate of AI’s contribution. It only credits what the tracking setup sees, and it gives no credit to an AI touch that came earlier in the journey.

What it does well

• It is auditable. A marketer can trace a conversion to a session, referrer and transaction or lead record.
• It is easy to establish. Most analytics platforms already offer a last-click or session-based view.
• It creates a consistent baseline. Teams can monitor recognised AI referrals over time, provided they keep their definitions stable.
• It is useful for operational questions. For example, it can show whether a known AI referral sent people who signed up or bought.

For e-commerce, connect the analytics conversion to the order record where possible. For B2B, define the conversion carefully: a demo form submission is not the same as a qualified opportunity or closed-won deal. A CRM connection can show whether an AI-referred session later became pipeline, but that connection only exists when identity and tracking data persist.

Where it fails

Last click cannot credit an interaction that did not create a recorded touch. If a person reads an AI answer, remembers the brand and searches for it in Google, last click may assign the visit to Organic Search. If the person types the domain, the session may appear as Direct. If an app or privacy feature removes the referrer, a real AI referral may also appear as Direct.

It also undervalues early research. In a purchase journey, a person might use an AI assistant to compare options, visit a review site later, and return directly to buy. Last click credits the final direct visit, not the initial research interaction.

Arcalea’s analysis of ChatGPT-assisted transactions describes this problem: in its reported B2B sample, a substantial portion of revenue associated with ChatGPT involved an earlier touch rather than the final recorded source. The study is a specific sample, not a universal rate, but it shows why last-click revenue should not be presented as the total value of AI discovery. Read Arcalea’s analysis of ChatGPT-assisted revenue.

Last click is also vulnerable to channel classification rules. Search Engine Journal reported Google Analytics adding an AI Assistant default channel group, with recognised assistant traffic eligible for separate classification. For practical configuration steps, Growth Rocket explains how to track ChatGPT traffic in GA4. However, channel rules do not make every AI visit identifiable: traffic from Google AI Overviews may remain within Organic Search, and an assistant not recognised by a configuration may be listed as Referral. Check your own property’s documentation and channel rules rather than assuming the label captures every engine. See Search Engine Journal’s coverage of the GA4 AI Assistant channel.

How to use it responsibly

Treat last-click AI revenue as confirmed, observed activity. Make the definition visible in reports: list the domains or channel rules included, the conversion event, and whether the number represents first session, last session or last non-direct touch.

Do not label it “total AI revenue”. A more accurate label is “revenue from recorded AI-referred sessions” or “AI referral conversions, last click”. Include a note that zero-click influence and unattributed visits are excluded.

Assisted and multi-touch attribution: the recorded journey

How it works

Multi-touch attribution (MTA) looks at more than the final session. A model may give credit to the first recorded touch, the last touch, or several interactions along the path. A linear model divides credit evenly across recorded interactions. A position-based model gives more weight to selected stages. A data-driven model uses observed patterns in the platform’s data to distribute credit. Pierview provides an overview of how teams approach AI search attribution across complex multi-step paths.

An assisted-conversion report is a simpler way to inspect whether a channel appeared earlier in a recorded path before another channel closed the conversion. The label and method vary by analytics tool, so check whether a report means “appeared anywhere before conversion” or calculates fractional revenue credit.

Imagine a buyer clicks a Perplexity citation, returns through Google Search, then comes back directly to purchase. Last click may credit Direct. An assisted report could show the earlier AI referral. An MTA model could allocate some credit to each recorded touch.

What it does well

• It shows sequence. Teams can see whether AI referral traffic tends to begin a journey, assist later research or close a conversion.
• It gives more context than last click. A recorded AI visit can be visible even when it is not the final session.
• It can support channel comparisons. Teams can compare the observed contribution of AI referrals with other recorded sources using one consistent model.
• It connects to downstream outcomes. With suitable analytics and CRM or commerce integrations, a first recorded visit can be associated with later revenue.

MTA is most helpful when buyers make repeat visits and the business has enough reliable first-party data to connect those visits. It can be useful for e-commerce products with a considered purchase, and for B2B journeys where a prospect researches, returns, then books a meeting.

Where it fails

MTA still depends on recorded touches. It cannot assign credit to an AI answer that did not generate a click, or to a click that tracking could not identify. Multiple touchpoints do not make invisible interactions visible.

Identity and privacy constraints can also break a journey. A buyer may research on one device and convert on another, use a browser that limits tracking, or return after a long delay. In each case, the system may not know that the sessions belong to the same person. A model can only distribute credit across the events it has connected.

A second risk is false precision. A model that assigns 20% of a sale to an AI referral may look exact, but the number depends on the model’s rules and the completeness of its data. Fractional credit is an allocation, not proof that the interaction caused that portion of the purchase.

How to use it responsibly

Use MTA to describe the recorded journey, not to claim that it captures the whole journey. State the attribution window, model type, channels included and identity limitations. Compare more than one model when the decision is important; if the result changes sharply between first-touch, last-touch and data-driven views, that sensitivity is itself useful information.

For a B2B team, carry a source field into the CRM and retain the original source where possible. Report both the initial known source and the most recent known source. Do not overwrite the first touch when a later visit occurs. For an e-commerce team, compare analytics sessions with order data and survey answers rather than assuming one platform has a complete customer history.

Self-reported attribution: what buyers remember

How it works

Self-reported attribution asks a customer how they heard about, discovered or chose your business. The question can appear on a post-purchase page, a lead form or during onboarding. It may use a list of choices, an open-text field, or both.

A response such as “ChatGPT recommended you” reveals an influence that may not appear in session analytics. The visitor may have arrived by searching for the brand or typing its URL. The survey does not prove the exact session or prompt, but it adds a useful piece of evidence about discovery.

Fairing’s LLM Product Discovery Benchmarks report that its e-commerce research found last-click tracking missed 74.5% of LLM product discovery in the brands studied. This is a provider’s benchmark, not a universal rate for every category or survey. Use it as evidence that the gap can be material, not as a correction factor to apply to your own analytics.

What it does well

• It can surface zero-click discovery. A buyer can name an AI assistant even if it never sent a tagged visit.
• It captures the buyer’s language. Responses can show whether people describe an assistant as a search tool, adviser or recommendation source.
• It is quick to start. A short question can be added to an existing form or post-purchase flow.
• It complements analytics. It can suggest where to investigate when Direct or branded search rises without a clear explanation.

Survey answers are especially useful when a business has a long or fragmented buying journey. A B2B prospect may use an AI assistant to make a shortlist, then arrive through a colleague’s link or a branded search. An e-commerce customer might see an AI-generated recommendation and later buy through a saved bookmark.

Where it fails

People may not remember every source. Some may say “Google” when they mean search generally, or name an AI assistant even if it played only a small role. Others may skip the question or choose the first answer that seems close. Responses can also be influenced by where and when the question is asked.

Surveys do not provide a reliable timestamp for every touch, a complete click path or a clean link from one answer to a specific AI citation. Open-text responses can be hard to categorise consistently. A single person’s answer should not be treated as a causal measurement of revenue. For addressing these unrecorded discovery paths, The Matchbox shares a zero-click dark funnel playbook that outlines practical tracking alternatives.

There is also a sampling issue. People who answer a post-purchase survey may differ from people who do not. If response rates vary by product, device or customer type, the results may overrepresent some groups.

How to set up a useful survey

Keep the question neutral and easy to answer. For example:

How did you first hear about us?

• Search engine
• AI assistant, such as ChatGPT, Perplexity or Claude
• Social media
• Friend or colleague
• Review site or publication
• Other
• Not sure

Then offer an optional follow-up: “Which AI assistant or source do you remember?” This captures detail without making every buyer type an answer. Fairing provides a guide to HDYHAU survey wording with examples for post-purchase attribution.

Avoid leading language such as “Did ChatGPT tell you about us?” That wording suggests a preferred answer. Do not ask several overlapping questions at checkout. For a B2B form, make the field optional and avoid adding friction to the primary conversion.

Review open-text answers regularly and use a consistent classification scheme. Keep “AI assistant” distinct from “search engine”, and preserve the original response for later review. If customers name specific products, do not assume every answer refers to the same type of use.

How to report it

Report the response count, the share of respondents naming AI, the collection point and the period. If you compare survey share with analytics sessions, label each measure separately. One is based on responses; the other is based on recorded visits. They have different denominators and should not be added together as if they represented unique conversions.

If Fairing or another survey provider reports a change over time, check that question wording, placement and answer choices stayed consistent. Fairing reports a more-than-tenfold increase in consumers explicitly naming an LLM between January 2025 and mid-2026 in its data. That trend is specific to its research and should not be assumed to apply at the same rate to every brand. It does, however, support testing a clear AI option in your own survey.

Incrementality testing: did AI visibility cause a lift?

How it works

Incrementality testing estimates whether an activity changed an outcome compared with what would likely have happened without it. Instead of following an individual click path, the analysis compares outcomes across a treated group and a credible control, or compares actual results with a modeled baseline.

For AI search, a team might increase useful content and answer-engine visibility work in one market while holding another similar market steady. It can then compare changes in qualified demand, branded search, direct traffic or sales. Another approach is to test an intervention against a pre-defined baseline and account for other changes that could affect the result.

The important idea is the counterfactual. A rise in AI mentions and a rise in sales at the same time do not prove that one caused the other. A test tries to estimate what sales might have been if the visibility work had not happened.

What it does well

• It addresses causality more directly. A well-designed test can estimate whether an intervention contributed to an outcome.
• It includes untracked influence. Aggregate outcomes can move even when individual AI journeys are invisible in analytics.
• It is useful for investment decisions. A credible lift estimate can help a leadership team decide whether to continue, adjust or expand a programme.
• It encourages a clear test plan. Teams must define the change, comparison group, outcome and time period before reading the result.

Incrementality is often the strongest option for an executive question such as “Did this programme increase qualified demand?” It is not a replacement for analytics. It works alongside observed conversion data and surveys, each of which offers a different view.

Where it fails

A test is only as good as its design. Markets may differ in brand awareness, seasonality, competition or media spend. A concurrent campaign or product change can affect results. If the intervention is applied everywhere at once, it can be difficult to find an unaffected control.

AI visibility is also hard to isolate. A content change can affect classic organic rankings, referrals and AI answers at the same time. If several parts of the marketing programme change together, a measured lift may reflect the combined work rather than AI citations alone.

Finally, experiments require enough data to distinguish a real effect from ordinary variation. A small business with few conversions may not be able to run a clean geo-test. In that case, a directional analysis can still inform decisions, but it should not be presented as conclusive causal proof.

A practical testing design

Start with a narrow question. For example: “Does publishing and promoting a set of comparison pages increase qualified non-brand demand in the test market?” This is more testable than “Does AEO work?”

Then define:

• The intervention: What content, technical change or distribution activity will happen, and where?
• The outcome: Which business measure matters, such as qualified opportunities or contribution margin?
• The comparison: Which market, audience or time-based baseline represents the likely outcome without the intervention?
• The guardrails: What else could change during the test, including paid spend, pricing, product availability or sales coverage?
• The visibility indicators: Which prompts, engines and citation patterns will be monitored to confirm the work was actually seen?

Track AI visibility as a diagnostic, not as the final outcome. A rise in citations can show that the intervention changed exposure. It does not, by itself, show that it caused revenue. SearchFIT’s overview of content strategies for AI search visibility can help teams plan the visibility work being measured.

Choosing the right model for your business

If you are an agency

Use last-click reporting as the baseline that clients can verify. Add an assisted view to show recorded AI referrals earlier in a journey, and include a consistent survey question where the client’s conversion flow allows it. For strategic programmes, propose a lift test only when the client has enough data and a suitable comparison.

If you are evaluating service delivery options for client reporting, compare our platform against manual tracking in SearchFIT vs agencies.

The key is to avoid presenting a single AI revenue figure as if all methods agree. Give clients a clear evidence ladder:

• Observed: recorded visits and conversions from recognised AI referrals.
• Reported: buyers who say an AI assistant influenced discovery.
• Modeled: estimated lift from a controlled or baseline comparison.

That makes the report useful without overstating certainty. If you are building repeatable reporting across clients, choose the same definitions and question wording where possible, while allowing for differences in conversion type and sales cycle.

If you are a B2B SaaS business

Connect source information to CRM stages. A demo request is an early conversion, not a closed sale. Preserve the first known source, latest known source and self-reported answer, then report how each relates to qualified pipeline and closed-won revenue.

This visibility matters as buying habits shift: industry research indicates that 42% of B2B buyers now use AI search to evaluate vendors. Use last click to identify observed referral sessions, MTA to inspect recorded research paths, and surveys to ask about AI-assisted vendor discovery. If AI visibility work is a major investment, consider a market or content test with clear controls. Keep lead quality in the outcome: more visits are not necessarily more sales-ready opportunities.

If you are an e-commerce brand

Pair storefront analytics with order data and a post-purchase question. Last click is useful for orders from known AI referrals. Multi-touch can show whether those visits appear earlier in recorded paths. A survey can reveal product discovery that ended in Direct or branded search.

For high-consideration products, buyers may research over a longer period or across devices. In that situation, survey responses and aggregate lift become more valuable because click tracking may fail to connect the whole journey. Compare reported AI discovery by category or product only if the response volume supports a meaningful comparison.

If you have limited data or resources

Begin with the methods that are inexpensive and easy to maintain: define AI-related referrals, check channel rules, and add a short, neutral survey question to a high-intent conversion point. Keep a regular record of visibility, branded demand and conversions. Do not buy a complex attribution tool before you know which decision it needs to improve.

A practical measurement stack can start with standard web analytics, a CRM or order system, a survey field and an AI visibility tracker. If you need custom dashboards across your warehouse and marketing sources, D23 offers managed Superset and plain-English AI analytics. SearchFIT’s pricing page outlines its platform options, while the right attribution setup will still depend on your analytics and business systems.

A comparison of effort, confidence and blind spots

The table is a decision aid, not a ranking. A business with a short buying cycle and well-tagged traffic may get strong operational value from last click and MTA. A business with substantial dark discovery may need surveys. A leadership team deciding whether to invest further may need an incrementality test.

Common mistakes when measuring AI search

Treating Direct traffic as proof of AI influence

A rise in Direct sessions can have many causes: returning customers, untagged campaigns, bookmarks, typed URLs or missing referrers. It may be consistent with AI-driven discovery, but it cannot identify AI as the source on its own. Use survey responses and broader evidence to investigate the pattern.

Treating all AI referrals as one consistent channel

Analytics rules differ, and they can change. Build a documented list of assistant referrers and review the channel classification in your own property. Keep Google AI Overviews separate conceptually from external assistant referrals: a citation click from a Google feature may be recorded as Organic Search rather than as a distinct AI channel. Search Engine Journal’s report on the GA4 channel is a useful starting point, but always validate the current setup in your own analytics.

Adding unlike numbers together

Do not add last-click revenue, survey-reported revenue and modeled incremental revenue. They may describe overlapping orders, use different populations and answer different questions. Present them as separate evidence types, with a note about how each was calculated.

Calling correlation causal proof

If AI visibility, branded searches and sales rise together, that is an interesting relationship. It does not prove AI visibility caused the sales increase. Other campaigns, seasonality, product changes or market conditions may explain some or all of the movement. Use a credible control or baseline when you need a causal estimate.

Chasing tool precision instead of data quality

A platform can produce detailed charts without seeing zero-click discovery or connecting sessions across devices. Before choosing a tool, ask what data it uses, how it identifies AI traffic, how it joins conversions, and what its model assumes. The tool name is less important than whether the output answers a decision you actually need to make.

A practical implementation plan

Establish the verified baseline

Document your conversion events and define which AI referrers or channel labels count as recognised AI traffic. Check how your analytics classifies ChatGPT, Perplexity, Claude, Gemini and other relevant sources. Keep the rules in a shared document and date any changes.

Record conversions and revenue from those sessions, then validate a sample against order or CRM records. If the analytics platform labels AI traffic differently from your own report, explain the mapping. This is your observable floor, not your estimate of all AI influence.

Add a buyer-reported signal

Add one neutral source question to a post-purchase page, lead form or onboarding flow. Include an AI assistant option and an optional follow-up naming the assistant. Record when the answer was collected and where it appeared.

Review responses for unclear categories and update the question only when needed. If you change the wording, mark the date so you do not compare unlike survey results without a caveat.

Inspect the recorded path

Use assisted-conversion reports or an MTA view to check whether AI-related sessions occur earlier in observed journeys. Compare first-touch, last-touch and, where appropriate, data-driven outputs. Note the attribution window and what the model cannot see.

For B2B, carry the data into the CRM and assess lead quality and opportunity outcomes. For e-commerce, compare sessions with orders and customer value. In both cases, do not confuse fractional credit with incremental value.

Run a test when the decision warrants it

Choose a specific intervention and outcome. If you have suitable markets or audiences, define a test and control before the work starts. If you do not, use a transparent baseline analysis and call it directional rather than causal.

Monitor visibility, citations and content changes as evidence that the intervention was delivered. Then assess whether the business outcome changed relative to the comparison. Record other marketing and product changes that could have affected the result.

Report evidence in separate layers

A clear executive report can contain three sections:

• Verified floor: conversions and revenue tied to recorded AI-related sessions.
• Buyer-reported discovery: survey answers, response count and collection method.
• Estimated incremental impact: test or modeled result, with the comparison and limitations stated.

This structure gives decision-makers a useful view without implying more certainty than the data supports. If the methods disagree, explain why: they may be measuring different parts of the journey rather than contradicting one another.

Verdict: which AI search attribution model should you use?

There is no universal winner. The best model depends on the question you need to answer.

• Choose last click when you need a simple, auditable baseline of conversions from recorded AI visits.
• Choose assisted or multi-touch reporting when you want to understand how known AI referrals appear across a trackable journey.
• Choose self-reported attribution when you need to uncover AI discovery that may have ended in Direct, branded search or another unlabelled visit.
• Choose incrementality testing when you need to estimate whether an AI visibility programme changed a business outcome enough to justify investment.

For most organisations, the practical choice is a combination: last click for the verified floor, multi-touch for observed paths, surveys for the dark funnel, and incrementality when the decision has enough value to justify a test. Keep the outputs separate, name the evidence behind each one, and do not claim that any single method captures all AI search influence.

For your next step, map the conversion and visibility metrics you already track, then use our resources to plan what to measure next.