← New run

Method

Data

• ListeningMind Data API path_finder returns a sample of search paths (query sequences, 300 by default) for the seed query. Per-path frequency is not provided.
• A path with the seed at the front is a journey after the seed; with the seed at the end it is an inbound path into the seed. Leak calculations use only journeys after the seed.
• keyword_info(data_type=all) attaches to every node its average monthly volume, trend, SERP features and 48 months of monthly volume (with gender and age mix for kr) [measured].
• The comparison point is a separate sample: the same seed queried again with time_point=3m/6m/12m.
• Social mention volume (Sometrend GetKeywordTransitions) is MCP-only and is not called from the server. Transition is computed from the ListeningMind time-point comparison.

Node classification

• Rule layer: if an alias of my brand or a competitor appears in the query, the node takes that brand. Words like recommendation, ranking, tier, value-for-money or forum names mark the compare stage; words like types, beginner or difference mark the explore stage.
• LLM layer [classified]: only nodes the rules leave undecided go to the LLM (Gemini, OpenAI as fallback), which judges brand presence, kind (brand/model/generic/comparison/community/off_category) and whether the query is inside the category. It is told never to invent a brand that is not in the query text.
• Channel layer: retailers, marketplaces, carriers, platforms and operating systems are classed as channels, not brands (kind=channel, no brand). The judgment is category-independent and has three layers. (1) The entity dictionary: a shipped list per market (seed) on which LLM judgments from runs (llm) and user corrections (user) accumulate; a category-scoped entry beats a global one and user beats llm. A query the rule layer left without a brand that contains a dictionary channel or platform name becomes a channel without going to the LLM. (2) During a run only newly found brand names are judged brand/channel/platform/other for that category, and the answer is written to the dictionary (channels and platforms globally too). (3) 'Filter channels and platforms', the per-brand 'Make category term' and node reassignments in the Evidence tab are stored as user entries and apply to every later run in any category. A declared competitor that looks like a channel is only flagged. The dictionary can be viewed and edited at /entities.
• Reassigning a node in the Evidence tab recomputes everything without credits.
• Input checks: before a run, the seed, my brand and the competitor list are checked against the entity dictionary and warnings are shown (a seed registered as a brand, a my-brand registered as a channel or platform, a category word inside the brand name, channels or platforms in the competitor list, my brand also listed as a competitor). No credits and no LLM.

AI Overview presence

• When the serp response carries an AI Overview block, each paragraph and list item is checked for my brand and the competitors (declared aliases plus brand names the classifier found). The brand named most is 'the brand recommended most'; ties go to the one named first [calc].
• Status: safe = my brand named most, tied = my brand equal with the leading competitor (competitor first), risk = a competitor named most, no brand = an AIO with no known brand, no AIO. A run without a my brand shows 'competitor' only.
• SERPs are fetched for the top 3 leak hubs at run time, and with the option 'Check AI Overview automatically' (on by default, 20 nodes by default) also for the nodes on marketer-view lines and the top leak hubs up to that count. The rest come from 'Check AIO' in the journey tree, the marketer view and the drill-down at 30 credits per node. Results are kept and shared by the tree dots, the path panel, the marketer view, the summary tile and the report.

What this tool measures and what it does not

• This tool finds the points in search-journey data where a brand gains or loses demand. It does not measure shares of consumers. Every share on screen is a share within the returned path sample.
• Observed Brand Reach = the share of sampled paths that first reach a query of that brand. Demand Exposure = the share of the summed average monthly volume of the brand queries reached (deduplicated). Neither is market share.
• Evidence level: passing paths n<5 very low · 5 to 9 low · 10 to 29 directional · 30+ strong. This is evidence maturity within the path_finder sample structure, not a statistical confidence interval; 0% and 100% must always be read together with n.
• Shrinkage: the leak strength used in the priority score is (competitor-reached paths + 4 × the run-wide competitor reach rate) / (brand-reached paths + 4), so smaller n is pulled toward the overall average. Raw shares on screen are left as they are, with n shown beside them.

Demand Capture Map

• Nodes that are not competitor-brand queries are split into four groups by the attribution of the journey after them. Winning = my brand holds 70% or more of demand exposure, Leaking = competitors hold 70% or more, Contested = in between, Uncaptured = no brand is reached.
• Leaking and Contested are ordered by priority score, Winning and Uncaptured by volume. Winning is what to keep; Uncaptured is demand no brand has connected to yet.

Marketer view (selecting lines to watch)

• Every branch of the journey tree (from the node after the seed) is a candidate, and rules pick up to 6. Types and conditions: Conquest = a non-brand node with 5+ passing paths, 40%+ of later paths reaching competitors, and my brand reached by 15% or less or downstream competitor demand at least 3× my demand. Ranked by downstream competitor demand × paths × (competitor reach − my reach). Up to 2.
• Capture = a non-brand node with 8+ passing paths and 50%+ of later paths reaching no brand. Ranked by paths × (1 + √volume). Up to 2. Defend = a branch where a my-brand node (3+ paths) has competitor nodes below it. Poach = a branch where a competitor node (5+ paths) is followed by 2+ generic or comparison queries. Strengthen = a non-brand node whose downstream my-brand demand is at least the competitor demand and my reach is 30% or more. One each.
• Concentration point: when 70% or more of a parent branch's demand sits in one child and that child keeps at least half of the parent's paths, the view steps down to that child. It stops before a brand node. Within one type a parent and its child are never both selected.
• The ad strategy frame (objective, queries to bid on, competitor queries, channels, KPIs) is a fixed rule per type, and every query is taken only from search terms that actually appear in this result's tree. The rule sentence and numbers under 'Why it matters' are computed; the sentences below them, the message angle and the caution are LLM narrative [narrative].

Leak definition

• Path attribution = the first brand node met in the journey after the seed: my for my brand, comp for a competitor, exit if an off-category query comes before any brand, none if no brand appears.
• Exit mix of node A = the attribution shares of the segment after A across paths passing through A [calc].
• Demand-weighted share = brand nodes reached after A, collected per brand without duplicates, summed by average monthly volume. Path-sample share counts paths. Both are shown; the priority score uses the demand-weighted one.
• leak_index = competitor demand-weighted share × brand reach rate (the share of paths that reach any brand after A).
• Brand reach rate = the attribution share of each brand across all journeys after the seed (by paths and by demand exposure).

Priority score

• Score (0 to 100) = (30 scale + 30 leak strength + 20 my-brand absence + 20 momentum) × evidence weight. Each component is a rank percentile among candidate nodes in the run. Evidence weight = n 30+ 1.0 · 10 to 29 0.85 · 5 to 9 0.65 · under 5 0.45. So a 100% leak on 4 paths does not outrank an 81% leak on 38 paths.
• Scale = average monthly volume [measured]. Leak strength = leak_index. My-brand absence = 1 − my demand-weighted share. Momentum = current leak_index − leak_index at the comparison point (nodes absent at the comparison point start from 0).
• Without a my-brand input the absence component is dropped; without a comparison point the momentum component is dropped; the rest is renormalized.
• Candidates = nodes that are not competitor-brand queries, not the seed, not off-category, and have 2 or more passing paths.

Funnel and transition

• The funnel (Sankey) folds each path into the columns entry → explore → compare → brand (first reached) → brand deepening. Moves back up the columns are not drawn, only counted.
• Transition = change in reach rate and demand exposure per brand (first order), plus new and lost leak edges and nodes whose leak_index rose or fell (second order). The overlap of the two samples (shared paths, shared nodes, Jaccard) is always shown alongside. The smaller the overlap, the more the change mixes volume changes, sample composition and limit truncation with real behavior change.
• The deep option pulls extra path samples starting from my-brand queries and the top competitor queries and computes the share that moves on to another brand after considering that one (exit rate after brand consideration).

Narrative

• Narrative sentences are written by the LLM from the evidence pack only (the numbers and query texts above). A sentence whose number is not in the pack is dropped whole [narrative].

→ Entity dictionary

Source labels

• [measured] Value returned by the ListeningMind Data API as-is
• [calc] Deterministic calculation from measured inputs (path sample, volumes)
• [classified] Label assigned by the LLM (brand, stage). Editable in the Evidence tab
• [narrative] Sentence written by the LLM; every number is checked against the evidence pack
• [no data] No data available for this value