How we compared keywords, buyer questions and AI searches

Avouch research · Study 1 · 16 October 2026 · Review draft

Read the study · Buyer-question register

The research question

Within the same client markets, how do saved SEO keyword lists, public buyer questions and recorded AI searches differ in wording and expressed buying concerns?

This is an exploratory comparison of existing records. It is not an experiment, a survey of all buyers or a measurement of missed market demand. Topic matching allows a more useful comparison than unrelated examples; it does not make the sources equivalent populations. In particular, the AI inputs were designed to test buying conditions, and the buyer-screening rules sought people with an identifiable need. Both choices influence the observed intent distributions.

The revision protocol was set after inspecting the data inventory and before the final comparative calculations. It was not externally preregistered. The analyst had seen earlier drafts and examples. The protocol records a later provenance correction that removed three keyword rows known to originate solely in AI searches. All final calculations were rerun after that correction.

The sources and dates

The live client store was read on 16 October 2026 at 11:54 UTC. Four client IDs were checked separately. Two supplied saved keyword data supporting these topics: financial services and data protection. The two other checked markets had no SEO rows in the examined collections. This is a statement about those checks, not proof that no other files or tools contain keyword research. Avouch’s own marketing-service cohort is excluded.

Search series contains 715 Google rank-check records, 633 Google query/page records, 114 Google AI page records, 2,344 Bing query records and 496 Bing site records. All tracked SEO seed terms were already in the union of keyword economics, targets and the client map. The wider series mixes engines, measurement types and dates; it was not pooled into a single keyword denominator. Bing queries were not analysed in this revision. Google query/page rows and AI page rows were not treated as additional independent query observations.

The buyer source is the frozen 15 October voice-corpus export. The two client libraries held 176 financial and 2,344 data-protection extracts. Forum-source filtering identified 72 financial extracts. For data protection, a further stored-text filter for GDPR, 27001, DPO or data protection reduced 970 forum extracts to 356. Grouping by stored URL gave 43 and 295 candidate discussions respectively. These are research-library counts, not unique buyers. Selected originals were checked on 16 October. Some posts are older, and some data-protection posts describe non-UK situations. Relative dates on source pages were not converted into a purported exact publication timeline.

The AI source is the frozen 15 October citation export and its existing response-selection register. The 64 selected responses used here span 18 August–15 October 2026. All are Gemini 2.5 Flash records labelled live, answer, and gemini_grounding with non-empty search_queries. The strings come from the engine’s recorded web-search metadata. They are not assistant-created reconstructions or consumer conversation logs. The retained original register includes source IDs, timestamps and model fields.

These datasets do not represent the same time period, geography, users or acquisition process. No temporal or population comparison is claimed.

Selection and the unit being counted

Four topics were fixed: adviser/retirement planning (F), equity release (E), personal protection insurance (L), and data protection/GDPR (D). The first three belong to one financial-services market. Topic rules and exclusions are saved in build.py and code_analyse.py. Relevant category terms, local variants and contextual terms are included; careers, unrelated services, named-brand navigation, political pension news and provider employment material are excluded. Employer group insurance is outside the personal-protection topic. The topic boundary omits relevant adjacent subjects such as inheritance-tax planning, mortgages as a standalone service and some privacy synonyms. It is a defined slice of the clients’ research, not the whole market.

SEO: union keyword economics, keyword targets and the existing client keyword map; remove duplicate case/punctuation variants within client. The initial topic pass produced 169 unique phrases. Manual review excluded 19 irrelevant/navigation rows. A further three rows were sourced solely from AI query imports, leaving 147 primary phrases. Those three are GDPR consultancy UK, gdpr consultant for small business, and what is life insurance. They remain in a separate sensitivity file. Known AI origins are excluded; incomplete historical provenance means this is not a guarantee that every remaining phrase arose independently of earlier AI work. Missing volume and absent phrases never mean zero demand.

Buyer: candidate URLs were ordered by SHA-256 of the fixed string study1-three-way-v1| plus the stored URL. Reddit host and comment suffix variants were grouped during screening. Select up to eight eligible verified discussions per financial topic and twelve for data protection, stopping at the target or exhaustion. The adviser group yielded three. Selection is exploratory within a previously curated corpus; hash ordering does not make it a representative buyer sample. The screening ledger records manual decisions, unavailable originals, duplicate discussions and candidates not pursued after targets. Some exclusions use a grouped reason rather than a separately adjudicated rationale.

Eligible authors appear to be considering or arranging a service, cover, family decision or business compliance need. Exclude promotion, advice replies, retrospective testimonials, career questions, consumer complaints and purely technical implementation questions. Business compliance questions may precede a decision to hire a provider. English-language units only are measured. Identity and customer status are not verified.

Use the first substantive opening question or request, with an explicit question-shaped title taking precedence. Fragment titles count. Greetings do not. One equity-release question omits its conversational lead-in and keeps the substantive question clause; this is recorded in the register. Where the short question refers to “this” or “one”, surrounding context determines the topic, but only the short words receive intent codes and a word count. Wider context is used for illustration only. This avoids comparing whole forum essays with brief keywords and also loses some buyer detail. One discussion is one observation, not a guarantee of one unique person.

AI: reuse the earlier fixed response selection, then retain relevant topics. In that selection, case and whitespace were normalised and town-only input variants were grouped into patterns. A fixed hash seed (study1-empirical-2026-10-15) selected one wording per pattern; the latest recorded response to that wording was retained, with record ID as a tie-breaker. Counties, substantive constraints and other wording were retained. The full town list and source hashes are in the original selection manifest. This revision selects 64 patterns and all 235 returned query entries. There are 234 distinct queries: one data-protection query repeats. Queries within a response are related observations. Search-free answers are excluded, so the study cannot estimate how often Gemini searches or fans out overall.

Measuring language and intent

Word counting uses the same Unicode expression for every primary text: [^\W_]+(?:[’'-][^\W_]+)*. Apostrophe and hyphen compounds stay together; slashes separate words. Median means the middle word count, averaging the two middle values when needed. The wording comparison uses one short buyer unit, one SEO phrase or one returned query entry. Exact overlap instead lowercases and joins alphanumeric runs, removing punctuation. It measures identical wording, not semantic coverage, retrieval success or equivalent intent.

The ledger also retains literal first-person pronouns and interrogative openings. “Near me” counts as first-person wording, and a title without a question word can still be a question. These mechanical fields are not measures of conversational quality and are not headline findings.

One AI analyst, Codex, assigned and reviewed all codes without blinding to source type. Rules plus explicit corrections are retained in code_analyse.py; buyer decisions are in buyers.py. This is a transparent interpretive reading, not independent human validation or a reliable measure of hidden psychological intent. No significance tests, confidence intervals or inter-coder agreement are claimed. Several codes can apply to one text; percentages do not sum to 100. Bare categories can remain uncoded.

N and I were added during clarification before final coding. Ambiguity remains: “good” signals trust without specifying a test; “fair fees” signals cost without naming an alternative; a complete sentence can carry circumstances absent from a short title. A second coder might reasonably change borderline counts. The clearest findings should be assessed against their actual row texts.

Complete primary results

All cells below are counts. Each denominator is the number of texts in that topic/source; the same text may appear in several concern columns.

Only two query entries exactly match a primary SEO phrase: financial advisor Poole and chartered financial planner Poole. Both are in adviser/retirement planning. One buyer unit, “Equity release?”, exactly matches a primary SEO phrase after normalisation. That finding cannot establish that a keyword list misses a topic or that a search engine cannot connect different wording.

Checks against an exaggerated conclusion

Location and repeated-query sensitivity. Remove a fixed list of towns, counties and UK country terms from SEO/AI text, then merge resulting identical phrases. Buyer units remain as observed. The exact list is in code_analyse.py. The table gives count / median words. It does not remove every possible geographic phrase, and it is not a new buyer sample.

Removing the one repeated AI query gives 234 distinct queries and does not change any primary topic median. Restoring the three known AI-origin keyword rows gives 150 phrases and four exact matching query entries, again without changing topic medians. The map-only financial SEO sensitivity contains 30 terms; the other client has no equivalent planning-map field, so that is not a comparable two-market baseline.

Observed Google queries. Apply the topic filter to the financial query export and data-protection query-dimension daily documents, merge identical client/query strings, then manually exclude career, unrelated news, navigation and consumer compensation material. The initial pass contains 410 phrases; 42 are excluded, leaving 368. The final register has 257 financial queries and 111 data-protection queries. Both have a median of four words. Repeated daily observations count once, and dates identify capture/export dates rather than asserting a common search window.

In data protection, 23 of 111 distinct phrases explicitly concern cost or terms, including ICO registration fees as well as consultancy prices. Five phrases exactly match saved research prompts. Removing them leaves 106 phrases, of which 22 concern cost/terms; the median remains four words. Exact matching cannot identify the person or system responsible for a search and cannot exclude other automation. These are observed query records, not authenticated buyer searches. The narrower cost/terms rule used for this cross-check is retained in google_check.py. It is not a complete recoding of all Google queries across the nine intent categories.

Search Console shows where a site appeared, is affected by its existing content/visibility, and omits some data. It is not all demand. Google also includes AI-feature appearances in Search Console reporting. Therefore these records cannot be labelled purely traditional-search queries. Google AI features documentation

Prompt design and fan-out. The designed prompts already contain many conditions and selection concerns, so their searches cannot prove that the same proportions arise in ordinary consumer use. The two report examples reproduce complete stored query arrays, one dated 15 October and one 31 August. Four of five insurance searches contain sick-pay wording; three of five equity-release searches omit equity release. There is no keyword-only control for either case and no measurement of recommendation or conversion outcomes. The public buyer posts are separate sources, matched by subject rather than linked to these responses.

Counterexamples. The saved keyword list already includes price, age, qualifications and comparison terms in some topics. Google query records contain detailed questions. Some buyer openings are only two or four words. These observations narrow the conclusion to gaps in the particular saved lists, rather than a claim that SEO only contains short category terms or cannot capture intent.

Volume provenance for the report examples

Volumes are saved monthly search estimates, not measured enquiries, recommendations or visits. Preserve the original date, location and provider for each example. Historical map values of unknown geography are not averaged with explicit UK estimates.

For life insurance, the older client map contains 301,000 with no geography field. For lifetime mortgage it contains 6,600. Neither value is used in the report. Restricted vs independent financial advice has a map value of zero but no verified geography/source settings; that is not used to claim zero UK demand. This study does not assign a volume to buyer questions or to the 235 AI query entries.

Reproducibility and review status

The numeric results, Google cross-check, and comparison ledger contain the complete public-facing row counts, codes, SEO phrases and AI searches. Buyer texts are paraphrased in the public ledger; exact measured units and source URLs are held privately and linked from the question register. Source IDs and raw client exports stay private. These redacted files accompany the published report. Private source records are not published.

The private evidence directory retains raw exports, extraction hashes, candidate sources, screening decisions, exclusions, selected rows, complete AI response records and all calculations. Raw-source hashes are in extraction-manifest.json; a separate verification manifest records the revision’s outputs and dependencies. Scripts are filed under the dated research folder in ~/code/scripts/research, with run order and dependencies in its README. No paid data collection or new model run was needed for this revision.

The numerical checks and editorial review are the same analyst’s work, not independent replication. Both charts were rendered and visually inspected. The public article includes text equivalents for the charts and a source register for reviewing the buyer questions.

The revised study supports a measurable difference in these collected texts. It does not establish all-market proportions, missed search volume, a causal benefit from conversational prompts/content, or improved brand recommendations. Human review of source selection and intent coding remains appropriate before publication.

Return to the study