A ranking report can be accurate and still give a business the wrong impression. A page may hold the same organic position while advertisements, maps, shopping units, or an AI answer change what appears above it. An answer engine may mention a company without linking to its website. A search-data provider may return a technically successful response that contains the wrong language or an incomplete feature block.
The measurement challenge is therefore larger than collecting positions. A useful search intelligence system must explain what it observed, under which conditions, and what the observation can reasonably tell us about customer discovery.
A search result is an observation with context
Consider the query “Polish accountant London”. A person searching on a phone in Ealing may have a different practical need from someone researching the same phrase on a desktop in Warsaw. Translating the phrase into Polish adds another variable rather than creating an equivalent experiment.
For a Poland–UK product, the measurement unit should be a query paired with a market, language, device, search surface, and collection time. The system should also record the requested location, any effective location returned by the provider, and whether the result came from a live request or a cache.
The following is a recommended observation record, not a universal API standard:
| Dimension | What to preserve | Why it matters |
|---|---|---|
| Query | Original text, spelling, and a separate normalized form | Over-normalization can erase meaningful differences |
| Market | Country and, where relevant, city or coordinates | Local demand cannot be represented by a country label alone |
| Language | Query language and requested interface language | These are separate measurement choices |
| Client | Device, viewport if captured, and session conditions | A mobile result is not a desktop result with a smaller label |
| Surface | Google organic, Bing organic, Maps, AI Overview, AI Mode, or another named product | These observations have different structures and meanings |
| Timing | Requested, collected, and received timestamps | Delivery speed and data freshness are different properties |
| Provenance | Provider, endpoint, request ID, parser version, and raw-response reference | An analyst must be able to trace a conclusion back to evidence |
This level of detail is achievable with commercial collection interfaces, although individual fields and targeting options vary. For example, DataForSEO documents separate Google organic, Maps, and Bing endpoints, with distinct retrieval methods. Endpoint existence establishes an integration route; it does not establish equivalent coverage or accuracy in every market. DataForSEO SERP endpoint catalogue.
The recommendation is to keep separate measurement panels for Polish queries in Poland, English queries in the UK, and Polish queries in the UK. An additional English-speaking country should become its own panel when the product actually serves it. “English-language search” is not one geographic market.
Measure the page around the ranking
Organic position should remain in the database, but it should sit beside a structured record of result features. At minimum, distinguish organic links, advertisements, local packs, featured snippets, People Also Ask, shopping results, and AI-generated blocks.
Do not force these into one ordering system. “Third organic result”, “third local business”, and “third citation in an answer” describe different things. Preserve the original provider fields and derive clearly named metrics from them.
For a local service business, a map appearance may deserve more attention than movement from organic position eight to seven. For a retailer, product units may change the commercial meaning of an otherwise stable ranking. These are reasons to investigate the full result composition, not reasons to assign invented click values to each feature.
People Also Ask also requires a collection policy. Questions visible initially are different from questions revealed after simulated clicks. An expanded tree can be useful for discovery, but it should not be presented as the page every searcher initially saw. Similarly, a collapsed AI block and a fully expanded answer should be distinguishable in stored observations.
Some AI content requires additional retrieval. SerpApi documents a dedicated AI Overview endpoint and a page-token mechanism for obtaining an overview associated with a search. An integration that assumes every answer is included in the initial organic response can silently undercount the feature. SerpApi Google AI Overview API.
AI visibility has several outcomes, not one rank
The first useful distinction is between an observed answer and the audience that may have seen it. A monitoring platform can repeatedly run a prompt and record the response. That does not reveal how many real users asked that question.
Four measures should be kept separate:
| Measure | Operational definition | What it does not prove |
|---|---|---|
| Brand mention rate | Valid sampled answers mentioning a resolved brand, divided by all valid sampled answers | Audience size, endorsement, or a website visit |
| Citation rate | Valid sampled answers with at least one explicit citation to the target domain, divided by all valid sampled answers | That readers clicked, or that the cited page caused the answer |
| Recommendation rate | Valid sampled answers explicitly recommending the brand under a documented classification rule | A purchase or a universally favourable brand perception |
| Attributed visits and outcomes | Recorded visits and conversions assigned to an identifiable referring source | All AI-influenced demand, including visits without usable referral information |
An optional fifth measure is share of citations among a declared set of competitors. Its denominator must be published. Counting citation occurrences, unique cited URLs, or answers containing a citation will produce different results.
Brand resolution needs particular care. A company name may also be an ordinary word, a location, or another organisation. Maintain an alias dictionary and review ambiguous matches. In Polish, inflected forms can complicate exact matching; in English, abbreviations can create false positives.
The recommendation is to retain the answer, visible links, prompt version, collection route, and timestamp. Label a measurement as a consumer-interface observation or an API experiment. Do not silently combine them.
Perplexity's current Search API documentation describes structured ranked web results and directs developers elsewhere for generated answers with citations. Consequently, a call to that search endpoint is evidence about its retrieval output, not a measurement of the answer a consumer would see in the Perplexity interface. Perplexity Search API documentation.
First-party AI reporting changed in 2026
Older measurement advice needs a date check. Google announced dedicated Search Generative AI performance reports on 3 June 2026 and subsequently stated that the insights had rolled out worldwide by 31 August. The announcement describes impressions, pages, countries, devices, and time breakdowns. It should not be interpreted as a promise of prompt-level reporting or a full AI conversion funnel. Google Search Central announcement.
The current help page states that the Search report includes AI Overviews and AI Mode, excludes Search Labs experiments, and may not appear where a site has insufficient impressions. It also documents aggregation and preliminary-data limitations. These details matter when comparing first-party impressions with a third-party sampling panel. Google's Generative AI performance report documentation.
Microsoft introduced an AI Performance public preview in Bing Webmaster Tools on 10 February 2026. Its announcement describes citation counts, cited pages, and sampled grounding queries across supported Microsoft and partner experiences. Microsoft explicitly distinguishes citation activity from ranking or placement within an answer. This is useful first-party evidence for the covered surfaces, not a census of every AI assistant. Bing Webmaster Blog.
The design implication is to store these reports alongside sampled observations, with their own definitions. Never manufacture a single “AI market share” number by adding Google's impressions, Bing's citations, and third-party prompt results.
An additional integration question remains open: a metric visible in a product interface is not automatically available through an existing public API or bulk export. Verify the exact supported retrieval route before promising automated ingestion.
Crawl access, search inclusion, and training are different controls
Technical access is worth checking before commissioning elaborate content experiments. For ChatGPT, OpenAI documents OAI-SearchBot for search and GPTBot for potential model-training use, with independent robots.txt settings. ChatGPT-User serves certain user-initiated requests and is not the automatic search crawler. These distinctions help diagnose access problems; allowing a crawler does not guarantee a citation. OpenAI crawler documentation.
Google's generative-search guide, updated on 10 July 2026, continues to emphasise established SEO fundamentals and describes a Search Console inclusion control. It does not prescribe a special schema or an extra AI text file as a route to success. Its discussion of retrieval and query fan-out also helps explain why an answer's sources need not match the visible organic results for the original wording. These are Google's documented statements about its own system, not universal rules for every engine. Google's generative-search optimisation guide.
The practical recommendation is to verify indexing eligibility, crawl access, substantive page content, and relevant product or business information before testing speculative GEO tactics.
Build an experiment that can detect its own weaknesses
A useful provider comparison begins with a fixed sample rather than a convenient keyword list. One proposed pilot uses 60 query–market pairs: 20 Polish queries in Poland, 20 English queries in the UK, and 20 Polish queries in the UK. Include local, informational, commercial, branded, and ambiguous queries within each group.
Testing both desktop and mobile on Google and Bing produces 240 measurement cells. Running these through two providers on three collection days produces 1,440 scheduled observations. Repeating 20 selected cells three additional times through both providers adds 120, for a total of 1,560 observations before retries or follow-up feature requests.
Use comparable collection windows and randomise provider order. Fix organic depth, language, requested location, cache policy, and feature settings. Run live-versus-live comparisons separately from scheduled-queue comparisons. A missing capability should be recorded as missing, not approximated without disclosure.
Review a stratified subset against contemporaneous browser evidence where collection is permitted. Human inspection is a reference check, not an infallible reconstruction of every user's result. The point is to identify wrong locales, missing blocks, malformed links, and systematic extraction mistakes.
Record repeatability within each provider before interpreting disagreements between providers. If a search changes substantially between two nearby observations from the same source, a cross-provider mismatch cannot automatically be attributed to poor extraction.
Make uncertainty visible in the report
The dashboard should distinguish “not present”, “not collected”, “unsupported”, “failed”, and “unknown”. Replacing all five with zero creates false trends.
A top-ten request that does not contain a domain should produce “not observed in the collected depth”, not an invented rank of eleven. An AI answer without a brand mention should remain in the denominator of mention rate. A failed answer collection should not, but its failure must appear in a separate coverage metric.
For repeated AI prompts, calculate rates within a stable prompt panel and show the number of valid observations. Repeats from the same prompt are related observations. If estimating uncertainty, resample by prompt or query group rather than treating every response as fully independent. Such intervals describe variation within the study design; they do not make a hand-picked panel representative of all search demand.
A trustworthy search report lets a reader answer three questions: what changed, how reliably it was observed, and which business decision the evidence supports. More observations are valuable only when those questions become easier to answer.
Part of the AutoMaSERP search intelligence series.
Read our editorial approach