How to Read Google AI Search Impressions Without Double-Counting

Short answer: Google's generative AI Search report describes impressions for supported AI features. Its data is included in the overall Web Search report. Compare the two only after aligning the property or page, dates, search type, filters, aggregation and data maturity; never add one total to the other as if they were separate channels.

A leadership dashboard shows 2,000 Google Web impressions and another 250 AI-feature impressions. Someone reports 2,250 impressions and concludes that AI produced 250 incremental prospects. Both conclusions are unsupported. The AI count is already included in the Web figure, and an impression is not a person, visit or enquiry.

This guide is for founders, growth marketers and analysts interpreting first-party AI-search reporting. Growthcraft Editorial prepared it with AI assistance; Akshay's personal review is not claimed. The framework, calculations and examples are editorial synthesis using synthetic numbers, not client results. Product reports change, so confirm the current interface and metric definitions before building a production dashboard.

Five decisions to get right

• Separate Google AI-feature impressions from AI-platform citations, referral visits and confirmed enquiries.
• Do not add Google AI-feature impressions to the overall Web total: Google's report says the AI data is included there.
• Match scope and aggregation before calculating a percentage; a sum of page rows is not automatically a property total.
• Keep unavailable, preliminary and rounded values distinct from exact zero.
• Compare complete equivalent periods and treat a post-edit change as an observation, not proof of causation.

Use the scope review worksheet to document the contract, the impression-share calculator to check conditional arithmetic, and the report-review prompt for a bounded handoff.

Why first-party reports now matter

Google announced dedicated Search generative AI performance reports on 3 June 2026 and later stated that the insights had rolled out worldwide by 31 August. Google says the Search report covers impressions in AI Overviews and AI Mode and offers page, country, device and date views. The report is not a complete view of ChatGPT, Perplexity, Bing or every AI answer. A site may not see the report if it lacks sufficient qualifying impressions. Google Search Central announcement; Google report documentation.

On 24 September 2026 Google announced a Web multimodal search-type filter for the Search and generative AI reports, covering image-led searches from sources such as Lens and image uploads. Compare like with like: a Web text view and a Web multimodal view describe different search types. This change is a reason to inspect report settings, not evidence that every business has meaningful multimodal traffic. Google's multimodal reporting announcement.

Microsoft's Bing AI Performance report is a separate source. On 13 October 2026 it announced a beta Topics Portfolio that groups grounding queries and offers approximate topic context when enough data is available. Microsoft says classifications may be imperfect and figures are directional. Its citation counts and topic estimates have different units and surfaces from Google's impression report. Microsoft Bing's Topics Portfolio announcement.

What exactly is an impression?

Google describes these impressions as instances in which a link to your site was shown in a supported generative AI feature. The report shows impressions, not a list of every buyer question or an independently verified recommendation. A visible link does not establish that the user noticed it, clicked it, read your page or became a customer.

By contrast, Microsoft describes a Bing citation as a source displayed in a supported AI-generated answer, and its grounding-query information as a sample of retrieval phrases. A grounding query is not necessarily the user's original wording. Bing's counts are useful in their own right, but they are not a Google impression count or a combined market share. Microsoft's AI Performance explanation.

Referral sessions are another measurement: they occur after a visit reaches an analytics system that records it. Consent, browser behavior, referrer handling and configuration can all limit observation. An email or WhatsApp click is an interaction, not proof a message was sent; a CRM record is not automatically a qualified project. Keep these measures and their denominators visible rather than rolling them into a single AI conversion rate.

The comparison contract

Before dividing two figures, write down the source, report name, property, exact URL if applicable, start and end dates, time zone, search type, country and device filters, aggregation level, extraction time, and whether the data is finalized. Both observations must describe the same scope. For Search Console, report dates generally use Pacific Time; align any analytics reporting window explicitly rather than assuming it uses the same day boundary.

Choose one of two valid levels for the proposed ratio: property-level totals from matching charts, or the same exact page under compatible page filters. Do not take the AI page table, sum its rows and divide by a property-level Web chart. Google explains that the chart is aggregated by property, while a page table is aggregated by page. Multiple pages can therefore be represented differently in the two views.

Keep unavailable values in a separate state. Google's documentation says that values displayed as unavailable can become zeros in exported data. Retain the visible source state or an explicit missingness field; a zero in a download alone may not be a measured zero. Record preliminary data as preliminary and refresh it before comparing complete periods. Never reverse-engineer hidden search queries or infer that a missing report means no visibility.

Worked example: the same 28-day population

Suppose a synthetic property has 2,000 exact Web Search impressions in a finalized 28-day period. Its AI-feature report has 250 exact impressions for the same property, dates, Web text search type and filters. Both numbers use property aggregation. The conditional share is 250 ÷ 2,000 × 100 = 12.5%. The Web total remains 2,000, not 2,250.

Suppose the site is shown twice in a result, once through a normal result and once through an AI feature. A property-level reporting rule can count the property once in the overall chart while the AI-feature observation is also included in that overall population. This is why the calculation describes a share of the reported Web impressions under the stated contract, not two mutually exclusive channels. Do not infer the number of users who saw both presentations.

Now suppose a page table contains two rows of 180 and 90 AI-feature impressions. Their sum is 270. It would be wrong to replace the property chart's 250 with 270 without a deduplication contract. A page-level comparison can still be useful if both numerator and denominator refer to the same canonical page and aligned filters. Label it as that page's share, not the site's share.

If the AI-feature report is unavailable, the ratio is unknown. If both exact comparable counts are zero, the ratio is undefined because the denominator is zero. If the AI count exceeds the supposed Web total, stop: first check filters, reporting dates, aggregation, search type, revisions and whether one figure was rounded. Do not clamp the answer to 100% or silently swap the inputs.

A deterministic check in JavaScript

The following dependency-free example accepts already normalized numeric counts and three human-supplied confirmations. It does not connect to Google, verify the scope or decide whether an exported zero is real. Run it in a modern JavaScript runtime; a production pipeline should store the actual scope fields and validation evidence alongside the result.

function googleAiShare(ai, web, sameScope, complete, exact) {
  if (sameScope !== true || complete !== true || exact !== true)
    throw new Error('Reporting contract is incomplete');
  if (![ai, web].every(x => Number.isSafeInteger(x) && x >= 0 && x <= 1000000000) || ai > web)
    throw new Error('Invalid or incompatible impression counts');
  return { ai, web, percent: web === 0 ? null : 100 * ai / web };
}
googleAiShare(250, 2000, true, true, true);
// { ai: 250, web: 2000, percent: 12.5 }
googleAiShare(0, 0, true, true, true);
// percent is null; it is not zero.

Test known answers, zero denominator, zero numerator, the 100% boundary, missing confirmations, negative, fractional, non-finite and impossible counts. The companion calculator adds input parsing, visible validation, reset and a JSON export. Software can reject impossible arithmetic, but a successful calculation cannot prove that the analyst selected compatible reports.

A reporting pipeline that preserves provenance

A useful internal model has a report-snapshot record and separate observations. Give each snapshot an opaque ID, source, property or canonical page, dates in source timezone, search type, filters, aggregation, extraction time and maturity state. Each observation stores a metric name, exact numeric value or null, unit and missingness reason. Keep raw exports in an approved private location; avoid sending query strings or customer details to a public prompt.

At ingestion, compare the two scope contracts field by field. A matching date label alone is insufficient if one export uses a page filter, another uses the whole property or one includes preliminary days. Add a validation result with explicit differences rather than overwriting either input. On refresh, keep the old snapshot ID and new extraction time so a changed number is not mistaken for a genuine business event.

When source systems use different reporting clocks, choose a documented common window and retain the original time zones. Do not manufacture a precise daily join between a Pacific Time Search report and an analytics property using another timezone. Use the actual overlap or compare larger aligned windows after accounting for reporting delay. Never attach a Search Console query to an individual CRM contact without a lawful, supported identity link.

Use Bing as a separate evidence lane

Bing's 13 October beta adds topic-level context built from grounding queries. Its topic size, site counts and query counts are approximate, and Microsoft says some smaller or low-presence sites may not see the view. Topic classification can be imperfect and branded competitor queries can appear as opportunities. Treat those as review candidates, not guaranteed audience demand or permission to make claims about a competitor.

Keep a row such as “Bing: 60 citations, one provider-defined window” separate from “Google: 250 AI-feature impressions, same 28-day Web window.” Even if the dates look identical, the systems can include different experiences, deduplication and counting units. Adding 60 to 250 produces a number without a meaningful common denominator. Neither source, alone, shows the actual number of enquiries.

Turn observations into useful work

Start with pages that are already appearing and confirm that they answer the reader's question, state the author's scope accurately, cite primary evidence and link to useful next steps. Then examine whether a valuable buyer or technical question lacks a clear destination. Update existing material before creating another near-duplicate page. A candidate opportunity from a report still needs a human review of intent, relevance and evidence.

Agree one measurement owner and a review cadence. Compare completed 7-day windows with completed 7-day windows, or 28-day windows with 28-day windows, using the same filters and noting seasonal or product changes. Record appearances, visits, email and WhatsApp clicks, and confirmed CRM enquiries separately. If traffic rises after a page edit, that sequence does not establish causation. A stronger causal question may need a deliberately designed experiment and enough observations.

For a practical measurement handoff, use the accompanying worksheet and calculator; for an editorial interpretation of related prompt samples, read the AI visibility dashboard guide. If you need support connecting content, analytics and CRM evidence, explore forward-deployed marketing.

Sources and limitations

Primary Google and Microsoft sources linked above were reviewed on 29 October 2026. All numeric examples are synthetic. Availability, coverage, UI labels, data thresholds and retention can change. No example claims a client's performance, a verified recommendation, a particular ranking, an incremental user or a future outcome.