AI search can influence a decision before someone reaches your website. The reporting mistake is to fill that invisible step with a convenient explanation. I would keep three questions separate: which visit can we observe, which influence does the buyer report, and which outcome can we attribute causally?
Evidence map: compare like with like
| Source / design | Sample and period | What it measures | Interpretation limit |
|---|---|---|---|
| Pew | 900 US adults; 68,879 searches; March 2025 | Click from Google results | Association by result-page condition |
| Adobe | >1T US retail visit measurement base | Conversion among referred visitors | March N and absolute rates not disclosed |
| Your acquisition register | Qualified events in a stated period | Observed referral source | Does not reveal pre-visit research |
| Your buyer accounts | Respondents / eligible population | Recalled influence | Nonresponse and recall bias |
| A proposed experiment | Predefined eligible population | Incremental outcome | Requires a credible counterfactual |
The click is a limited observation
In Pew Research Center’s March 2025 US panel, visits to Google pages with an AI summary led to a traditional-result click 8% of the time, compared with 15% without a summary. A link inside the summary was clicked in 1% of visits with a summary. The analysis covered 900 adults and 68,879 unique searches; result pages were reconstructed the following month. These are observational differences, not a randomized estimate of the summary’s effect. Pew Research Center, Google users and AI summaries, July 22, 2025
The study makes the measurement boundary tangible: a visible answer and a website visit are different events. It does not show that those unanswered clicks became sales, that all non-clicking users were persuaded, or that your brand was present in their answers. Those questions need other evidence.
Direct traffic does not identify an AI source
Google Analytics defines direct / none as traffic without a clear referral source. Its documentation lists missing traffic information, direct URL entry, offline documents and ad blockers among possible explanations. An increase in that bucket cannot identify which explanation applies to a particular visit. Google Analytics, Understand direct / none traffic
When an assistant sends a visit with a usable referrer or campaign parameter, retain it. Check the landing URL and redirect chain, and verify that your analytics implementation preserves the information. A visit carrying an identifiable AI source belongs in an observed referral report. A visit without it remains unattributed unless additional evidence supports a narrower account.
Google also reports AI Overviews and AI Mode within the Web search type in Search Console. That aggregate is useful for search performance, but it is not a clean record of every AI exposure or an assistant-by-assistant acquisition report. Google Search Central, AI features and your website
Use three evidence layers
The first layer is observed acquisition: a recorded source, visit and qualified action. Report session scope and the conversion definition. A session attributed to an assistant shows an identifiable arrival under your tracking rules; it does not capture all earlier influence.
The second layer is reported influence: the buyer says an assistant helped them understand or compare options. Preserve the response, collection date and question asked. Allow multiple influences and an unknown answer. Report the number of respondents and nonrespondents so a small enthusiastic subset does not become a claim about all buyers.
The third layer is incremental effect: a credible experiment estimates what changed because of an intervention. Define treatment, comparison, outcome and time window before inspecting results. Without a credible counterfactual, use the word association. Adding the first two layers together does not create the third.
A worked example without invented attribution
Consider a hypothetical month with 20 qualified inquiries. Four have an identifiable assistant referral, six respondents say an assistant influenced their research, and two of those respondents are also among the four referrals. The observed union is eight inquiries, or 40% of 20, with at least one of these recorded signals. It is neither ten independent inquiries nor 40% of revenue caused by AI.
That example illustrates the data rule, not a finding about my site or a client. Join records only where the person has appropriately provided an identifier and the collection is permitted. Otherwise report aggregate layers separately. Do not guess that anonymous visits and questionnaire answers belong to the same buyer.
Make the next review defensible
Ask the analytics owner to audit source retention using controlled test visits through the actual link and redirect path. Ask revenue operations to define a qualified inquiry and manage the optional research question. Ask the research owner to retain the exposure panel separately. Keep failed tests and missing responses visible.
Start with existing tools. A new dashboard cannot reconstruct a conversation it never observed. If traffic-source capture is broken, repair it before using source changes to recommend a budget shift. If buyer responses are too sparse, improve collection and report the limitation rather than allocating unknown influence.
Review the layers alongside qualified pipeline, sales cycle and revenue. A rise in mentions may justify examining consideration; a rise in measured referrals may justify improving a landing page. Neither alone establishes incremental return. For a larger investment decision, commission an experiment that fits the sales cycle and addresses other changes in demand.
The next move
Bring one month of data to the next review in three columns: observed acquisition, reported influence and estimated incremental effect. Leave unavailable evidence marked unknown. Your team can act on partial evidence without claiming complete attribution.
Conversion quality is a different research question
Adobe’s April 2026 report draws on a measurement base exceeding one trillion visits to US retail sites. AI referrals converted 38% below non-AI traffic in March 2025 and 42% above it in March 2026. The summary does not disclose March subset sizes or absolute conversion rates. These are observed channel comparisons, not a randomized estimate of additional sales. Adobe, 2026
The chart makes the comparison readable by setting non-AI conversion to 100 separately in each period. AI therefore has an index of 62 in March 2025 and 142 in March 2026. This normalization is my arithmetic from Adobe’s relative figures. A value of 142 does not mean a 142% conversion rate, and the difference is not a 42-percentage-point lift.
The reference resets in each period. You cannot infer the change in the absolute non-AI rate, the number of orders or the revenue contribution from these indices. A channel can send relatively high-intent visitors while contributing a small share of total visits. The study does not disclose your company’s channel mix.
Visitor selection is central to interpretation. Someone who follows an assistant’s link after asking detailed comparison questions may already be close to buying. Product mix, device mix and campaign exposure can also differ across channels. A higher observed rate is a reason to investigate referral quality, not proof that the assistant caused the purchase.
Pew’s click observations and Adobe’s conversion observations address different stages. The first concerns clicks from Google result pages in one panel; the second concerns what referred retail visitors do after arrival. Fewer outgoing clicks and stronger conversion among the people who arrive can coexist. Pooling the percentages would erase the denominator that gives each result meaning.
Calculate the effect size you actually observed
Using Pew’s published percentages, 8% versus 15% is a difference of minus seven percentage points. Relative to 15%, it is approximately 46.7% lower. Neither calculation corrects for differences between queries with and without summaries. Label this a descriptive association, not an estimated causal reduction caused by AI.
For your own reporting, publish the numerator and denominator beside every rate. “Four qualified leads from assistant referrals out of twenty qualified leads” is interpretable. “AI drove 20%” leaves the reader guessing whether the measure concerns sessions, leads, orders, revenue or an experiment. Choose the business event first, then name the metric.
Keep raw counts when volumes are small. One additional conversion can move a channel percentage substantially. Report the period and any qualification lag so an incomplete recent cohort is not compared with a fully matured older cohort. A lead created yesterday may not yet have reached the stage you are measuring.
A minimum data dictionary for a credible report
Define qualified before examining the result: for example, a lead that meets documented service-fit criteria and has completed the agreed qualification step. That is a hypothetical definition, not the rule every business must adopt. Keep it consistent across channels or explain a change.
Use separate fields for observed referral source, landing page, event date, qualification date, unique opportunity identifier, reported research sources and missing-response status. The opportunity identifier supports deduplication; it should not expose personal details in a shared editorial report. Preserve the attribution rule and reporting window as metadata.
A self-reported question should allow multiple research sources and “do not remember.” Record how many eligible people were asked, how many answered and how many named an assistant. Six mentions among twenty opportunities are not the same measure as six mentions among eight respondents. Missing answers are unknown, not evidence of no AI influence.
For the existing hypothetical example of twenty qualified opportunities, four observed referrals and six self-reported mentions with two overlapping opportunities produce eight distinct associated opportunities. If only eight people answered the research question, also report six out of eight respondents and response coverage of eight out of twenty. The union of eight remains an association count; it is not eight incremental sales.
Validate source collection with a documented test visit and a test conversion where your analytics setup permits it. Check whether redirects, consent conditions and cross-domain navigation preserve the information you expect. Label test events and exclude them from commercial reporting. A technically intact pipeline can still miss research that happened before the visit.
Use an experiment only when the business question requires it
A decision to correct an inaccurate product claim can proceed from a verified factual error. A decision to claim incremental revenue needs a credible counterfactual. Specify the intervention, eligible population, outcome, comparison and measurement window before interpreting results. Where feasible, random allocation reduces selection bias; operational feasibility and contamination still need review.
Do not use an exposed versus unexposed comparison when exposure is merely whether someone chose to use an assistant. The choice itself may reflect interest or intent. Publishing a page and observing better conversions later also leaves seasonality, price changes and other marketing activity unresolved.
If a defensible experiment is unavailable, retain the observational label and make a proportionate decision. You may improve evidence, monitor qualified referrals and collect buyer context without claiming causal revenue. The useful report states what the team can change, who owns it and which later observation would strengthen or weaken the working explanation.
Sources and scope
Public-source analysis, checked October 12, 2026. Proposed frameworks and examples are not findings from an original study. US data is not extrapolated to Spain.
FA