Website analytics begins at the measured door. It can report how a visit was classified, which selected actions occurred and which configured event received credit. It cannot replay everything the person read, asked, compared or believed before the session began.
That missing context is the hidden research stage. It may include ordinary search results, AI-generated answers, review platforms, competitor websites, private recommendations, podcasts, social media and conversations inside the buyer’s organisation. Some encounters leave partial traces. Many do not create a clean buyer-level record you can recover later. Here is the line that keeps the analysis honest:
A missing record is not proof that nothing happened, but it is not permission to invent what happened.
I treat the hidden research stage as an evidence gap, not a ghost funnel. The job is to separate what systems observed, what those observations reasonably suggest and what must remain unknown.
TL;DR
- Analytics and Search Console answer narrower questions than most growth presentations admit; know each product's scope.
- A captured AI answer is evidence of the screenshot, not of a buyer.
- Classify every signal by what it can and cannot support before interpreting it, and keep the denominator.
- Report four columns, observed, scope, interpretation and unknown, then act only where evidence earns the action.
Dashboard theology fails in both directions
Dashboard theology is the belief that an event did not matter unless a platform plotted it. That is nonsense. The opposite superstition is equally weak: one favourable chatbot screenshot proves an invisible customer acquisition journey is working. Also nonsense.
Google Analytics uses different scopes for acquisition. Its User acquisition report describes how new users first arrived, while Traffic acquisition describes the source of sessions. Google notes that metrics can differ between those reports because their scopes differ (Google Analytics acquisition guidance). Neither report claims to reconstruct the thoughts, comparisons or unmeasured encounters preceding arrival.
A session attributed to Organic Search supports an attributed arrival under the property’s configuration. It does not prove that organic search created the need, introduced the brand or supplied the decisive argument. A direct visit does not prove that no earlier influence occurred. It means the recorded session entered through the route Analytics classified as direct. The reports are not lying. They are answering narrower questions than many growth marketing presentations admit.
Google Search leaves real but bounded gaps
Search Console supplies useful observations about Google Search: impressions, clicks, queries, pages and search appearances. It does not produce a diary of every person’s research. Google now provides a dedicated Generative AI performance report for link impressions from AI Overviews and AI Mode. The report can group impressions by page, country, date and device, while those impressions remain included in wider Web performance data (Google Generative AI report). It does not provide the exact question, generated wording, unlinked brand mentions, clicks, buyer response or separate performance for the two supported features. A dedicated impression view is useful. It is not a replay of the buying journey.
The ordinary query table also has boundaries. Google omits some queries for privacy, stores and displays a subset of rows, and notes that table totals can differ from chart totals (Google Search Console dimensions). Google explains that Search Console and Analytics can differ because the products process and scope data differently (Google explanation of Search Console data). Those limitations are not reasons to discard the product. They are reasons to stop treating an export as complete demand evidence.
Google Analytics also defines an AI Assistant channel for recognised arrivals from sources including ChatGPT, Gemini, DeepSeek, Copilot and Grok. Google AI Overviews and AI Mode are excluded from that channel and sit within Organic Search where relevant (Google Analytics channel definitions). The honest conclusion is not AI traffic is invisible. It is that available products describe recorded exposures and arrivals using product-specific groupings with known omissions.
An answer sample is not a buyer
Google says AI Overviews and AI Mode may use query fan-out across subtopics and sources while producing responses, and that links and answers can vary because different models and techniques may be involved (Google AI features guidance). ChatGPT Search can show citations and a Sources panel. OpenAI also warns that search results and citations can be incomplete, outdated or incorrect and advises checking cited material when accuracy matters (OpenAI ChatGPT Search guidance). Those facts justify inspecting answer surfaces. They do not make one captured answer representative of a market.
A record can establish that a named product returned a response, under documented conditions, for an exact question, at a stated time, with the visible wording and sources preserved. It cannot establish how many buyers saw the answer, whether a prospect believed it, why someone did not click, whether a competitor gained an enquiry or whether other users saw the same output. The screenshot is evidence of the screenshot. Nothing else arrives free with it.
Separate the signal from the story you want
Classify every item before interpreting it.
| Evidence | What it can support | What it cannot support alone |
|---|---|---|
| Analytics session or event | A measured arrival or action under the property’s configuration and attribution rules | Complete pre-visit research or motive |
| Search Console data | Recorded Google Search exposure and visits within report scope | Every query, generated wording, unlinked exposure or reason for action |
| Captured search or AI answer | Output and visible sources under recorded conditions | Audience reach, buyer exposure, causation or market share |
| Buyer’s voluntary account | What that person reports about their journey | How common the journey is across all buyers |
| Sales notes containing repeated questions | A recurring issue within the reviewed notes | Which channel created it or how many silent buyers shared it |
| Branded search, direct visit or later enquiry | A downstream action | Which earlier encounter caused it without another defensible connection |
This table is deliberately unsatisfying. Good. Marketing measurement becomes dangerous when uncertainty is polished into certainty because the board wants a cleaner arrow. My rule is to preserve the observation, attach its limitation and earn the next investigation. Do not smuggle a buyer biography into the space between two metrics.
Build a research record around a decision
Are we visible in AI search? is too vague to direct useful work. A better question is: when this named product answers these approved pre-purchase questions, how often does it describe our service accurately and provide a checkable source?
Freeze a small question panel tied to real commercial decisions: category, suitability, comparison, limitation, cost basis or implementation. Keep the exact wording. Do not replace an awkward result with a friendlier prompt halfway through the run. For every attempt, record:
Keep the denominator. Three favourable appearances mean little if a presentation hides seven absent, failed or incomparable attempts. Repeat the same bounded panel at a cadence justified by a decision. A launch might warrant closer temporary observation. A stable category may not. Monitoring unable to trigger investigation or action is anxiety with admin rights.
Ask buyers without putting the answer in their mouths
Sales teams often hear context analytics misses. Prospects mention a comparison, repeat an unfamiliar category label, name a competitor or arrive with a confident misconception. That evidence becomes useful when recorded rather than embellished. Add one neutral question to an appropriate enquiry or discovery process: what had you already looked at before contacting us?
Let the buyer answer in their own words. Do not ask Was it ChatGPT? and then celebrate when the suggestion produces agreement. Preserve the response where normal consent, privacy and retention rules permit. Separate the person’s report from the salesperson’s interpretation. Code themes only when enough comparable records exist to make the exercise useful.
I would rather work from ten properly captured statements than one hundred anecdotes beginning sales keep saying. The ten still do not become population-level attribution. They remain customer-reported evidence with an explicit denominator. The larger opportunity is often not a channel claim. It is discovering the question the website design repeatedly fails to answer before the conversation begins.
Being asked to prove where buyers came from before they arrived? We build the bounded research record and the four-column report so the unknowns stay visible.
Triangulation earns a test, not a verdict
Independent signal classes can point towards the same problem. Even then, the conclusion must remain narrower than the story a marketing deck wants.
| Observed pattern | Defensible response | Claim to reject |
|---|---|---|
| A recorded answer repeatedly describes the service incorrectly and the relevant page contains conflicting language | Correct the public definition and rerun the bounded sample | “The answer is costing leads” |
| Buyers report using a named product and their misconception matches a captured output | Investigate the shared misconception and improve source material | “The product caused every misunderstanding” |
| Search clicks fall while rankings appear stable and generative impressions shift | Inspect query mix, presentation, demand, pages and answer surfaces | “AI Overviews caused the traffic loss” |
| Branded demand rises after a campaign while some AI answers mention the brand | Record both and seek campaign and buyer evidence | “AI generated the brand demand” |
| A competitor appears in a small sample and later wins a deal | Review deal evidence and public comparison independently | “The citation won the deal” |
Correlated observations can justify a page review, message test or better measurement design. They cannot manufacture a causal chain. I prefer the smaller diagnosis that survives challenge. It usually points to a more precise repair too.
Make the unknown visible in the report
A serious pre-visit research report needs four columns.
Observed
The exact metric, answer capture or buyer statement. For example: the brand appeared in four of twelve usable responses across one fixed panel during August; two used an outdated service description; three further attempts produced no comparable answer.
Scope
Product, question set, market, period, sample size and configuration. The denominator lives here.
Interpretation
The bounded conclusion the observation supports. For example: the recorded product sometimes exposed the brand, and the obsolete description warrants a source review.
Unknown
The attractive claim the evidence cannot establish. For example: audience exposure, buyer influence, market visibility and the reason each response was produced.
That is less exciting than an AI share of voice score. It is much harder to misuse. Do not combine mentions, citations, accurate descriptions, recommendations, clicks and qualified enquiries into one percentage. The units have different denominators. A brand can be cited and described badly. It can be mentioned without a link. It can receive a visit producing no suitable opportunity. AI marketing becomes more credible when the objects stay separate long enough to teach the business something.
Join research to action through a decision gate
Before acting, complete this short gate:
- Which exact observation triggered investigation?
- Which independent evidence supports the same problem?
- What is the narrowest defensible interpretation?
- Which public page, source or business process owns the correction?
- What change can be made without inventing causation?
- Which result would show that the correction shipped?
- Which external outcome will remain unpromised?
The action might be correcting a service definition, publishing a material limitation, improving comparison evidence, fixing an inaccessible page or changing the sales question. It might also be no change. Do not rewrite a sound page because one output omitted the brand. Do not blame an answer surface for lost demand while seasonality, offer fit, tracking, competition and the page itself remain unchecked.
This is the difference between AI readiness and AI theatre. Readiness creates maintained facts, observable tests and decision ownership. Theatre fills every unknown with the newest channel name.
Stop when the evidence cannot carry another claim
The hidden research stage is not permission to monitor everything indefinitely. Stop when the investigation produces one of three outcomes: a specific evidenced correction; a better-controlled observation to run; or an unknown that cannot be resolved proportionately with available access.
Do not sample until randomness supplies the screenshot you wanted. Do not keep an expensive monitoring programme because the report can always generate another chart. Do not convert a buyer’s recollection into precise multi-touch attribution. There is an unavoidable boundary: the hidden stage cannot be reconstructed completely. People forget, research privately, cross devices, ask colleagues and return through different routes. Better instrumentation narrows some gaps while privacy, platform and human-memory limits preserve others.
Attribution does not need to recreate the buyer’s entire mind. It needs enough clean evidence to improve the next decision and enough visible uncertainty to prevent false confidence. Measure arrivals. Inspect search exposure. Sample answer products under recorded conditions. Capture voluntary buyer accounts. Compare independent signals. Put unknowns beside findings. Then act only where evidence earns the action. Stop filling the blank space before the session with a funnel diagram.
Good to know
FAQs
No. It means the recorded session entered through the route Analytics classified as direct. Acquisition reports describe attributed arrivals under the property’s configuration; they do not reconstruct the research that preceded the visit.
It provides a Generative AI performance report for link impressions from those features, grouped by page, country, date and device. It does not show the exact question, generated wording, unlinked mentions, clicks or separate performance per feature.
It is evidence that a named product returned that response under recorded conditions. It cannot establish how many buyers saw it, whether anyone believed it, or whether a competitor gained an enquiry.
Four columns: what was observed, the scope and sample size, the narrowest interpretation the observation supports, and the attractive claim the evidence cannot establish. Keep mentions, citations, clicks and enquiries as separate objects with their own denominators.
Next step
Want measurement that survives a sceptical question?
We separate what your analytics, Search Console, answer samples and sales notes can each support, then build a bounded research record and a report that keeps the unknowns visible.