Methodology and capabilities

What EllipseSearch measures, how each number is defined, and what the product does and doesn't do yet. Methodology version 2026.09.27 · capability list 2026-09-27.2. Reports record the methodology version they were computed with, so older reports can be read against the rules in force at the time.

What the product does today

Answer collectionAvailable
Answers are collected from ChatGPT, Perplexity, Gemini and Google AI Mode through Bright Data's live-product scrapers.
Limits: Each brand collects only from the engines enabled in its settings and covered by its schedule.
Daily monitoringAvailable
Paid plans run each tracked prompt once a day on the brand's enabled engines and markets.
Limits: What actually runs depends on each brand's schedule, shown as coverage (for example “Daily · 1/6 prompts · 1/4 engines”).
In-country collectionAvailable with limits
Every run records the country it was requested from, and the provider's reported egress country where it reports one.
Limits: Verified where available: some markets can't be collected in-country, and older runs have no location metadata. Unverified runs are labelled, never presented as confirmed in-country.
Arabic analysisAvailable with limits
Arabic prompts, Arabic-aware brand and competitor detection, and English/Arabic comparisons on matched prompt pairs.
Limits: Dialect, Arabizi and sentiment accuracy haven't yet been benchmarked with native-speaking reviewers. Language comparisons need matched prompt pairs.
Content characteristicsAvailable
Structure, data density and directness of answers and sources, scored by our analysis.
Limits: Inferred characteristics, not the engines' ranking factors or proof of why an engine chose a source.
Accuracy checksAvailable with limits
Claims about the brand are checked against confirmed facts and the latest trusted crawl.
Limits: Only claims that could be checked are reported. With no confirmed fact base, results say what was checked rather than calling an answer accurate.
Outcome reviewsAvailable
Legacy before/after reviews of a recommendation's linked prompts (records made before Track), using an exact two-sided test, one answer per prompt and day, and unchanged control prompts.
Limits: New changes are measured by Track's fixed-window policy instead. Results are observations, not proof that the change caused them.
Track: change recordsAvailable
A ledger of website edits, technical fixes, fact corrections, outreach and published placements with their dates, evidence, page checks and audit history.
Limits: Page checks verify implementation (the page shows the recorded text), not AI effectiveness. Attachments are type-checked and stored privately but not malware-scanned yet.
Track: outcome evaluationAvailable with limits
Fixed 28-day before/after windows around a 7-day lag, frozen equal-weight prompt groups, coverage and persistence gates, overlap and comparator guards.
Limits: Needs daily scheduled collection covering every group. Labels start as shadow evaluations for owners and admins until released. Descriptive, not causal: no p-values.
Track: published-placement monitoringAvailable
Third-party pages are checked for the brand passage daily for 14 days, then weekly; removal needs two conclusive checks at least 24 hours apart.
Limits: Checks respect robots.txt and never log in, so blocked pages show as access issues rather than removals.
Track: business-driver associationsAvailable with limits
Answers are tagged with evidence-grounded brand–driver associations (value, reliability, service…) and shown as a frequency and rate heatmap with passages.
Limits: Beta: a human-reviewed evaluation set hasn't approved the extractor yet, so generated insights are off and it can't be an outcome's primary metric.
WriterAvailable
Drafts grounded in the brand's confirmed facts and cited sources, with checks that block unsupported claims before export.
Limits: Offers and capabilities come only from confirmed facts; missing facts are marked for you to fill in.
Client reportsAvailable
Branded reports with a frozen snapshot of the data, formulas and recommendations at generation time.
Teammates and rolesAvailable with limits
Invite teammates as Owner, Admin, Editor or Viewer. Roles are enforced on the server and in the database.
Limits: Access is per workspace. Per-client access, comments and assignment notifications aren't available yet.
Email digests and alertsNot available yet
Weekly digests and alerts for drops and failed runs.
Limits: No notification emails are sent yet. Check the Activity tab for failed runs.
Traffic and conversion measurementNot available yet
Analytics, CRM and crawler-log integrations.
Limits: Not connected: EllipseSearch measures mentions and links in answers, not clicks, visits or revenue.
Prompt demandNot available yet
How often buyers ask a question in AI engines.
Limits: Engines don't publish prompt volumes; suggested prompts aren't ranked by demand.
Two-factor sign-in and SSONot available yet
Password sign-in with reset is available.
Limits: Two-factor authentication, session management and single sign-on aren't available yet.

Metric definitions

Visibility (mention rate): Completed answers that name the brand ÷ completed answers. Failed collections are listed separately and never count as "not mentioned". With no answers the metric is "not measured", never 0%. Fewer than 10 answers are shown as descriptive only, with a 95% range.
Share of voice: Answers naming the brand ÷ answer mentions of the brand and every tracked competitor. Counts names in answers, once per entity per answer. Cited websites never enter it. Associations, regulators and anything marked “not a competitor” are excluded.
Citation share: A domain's citations ÷ all qualified citations. A citation counts once per page per answer. Tracking parameters are ignored when matching pages. Placeholder, invalid and map-attribution records are shown separately and excluded (source extraction version 3).
Competitor-owned share: Citations of competitors' own websites ÷ all qualified citations. A site is a competitor's when its domain spells the competitor's name (or the distinctive part of it), or the cited page's title names the competitor and the domain resembles it. Competitors are those in the brand's settings that count in share of voice plus rivals named in answers. A directory or article that lists a competitor is not the competitor's site. A person's decision on a domain overrides the rules.
Authoritative share: Citations of official, institutional and editorial sites ÷ all qualified citations. Official: government and public-sector domains. Institutional: universities, encyclopedias, standards bodies and research publishers. Editorial: recognised news and trade publishers. This is our classification from the domain, not an engine's trust score.
No error in checked claims: Checked brand descriptions with no contradicted claim ÷ descriptions with a checkable claim. Only claims that could be checked against confirmed facts count. It is not a verdict on the whole answer.
Diagnostic score: Composite of mention, context, attribution and position (0–100). An EllipseSearch diagnostic, not an engine's assessment. A brand absent from the answer scores 20.

Sampling and outcome reviews

Each tracked prompt is collected at most once a day per engine and market on scheduled runs. Days are counted in Dubai time (UTC+4).
On scheduled runs, an identical request already collected for another brand in the previous 20 hours reuses that answer. Reused answers are marked and count once.
Outcome reviews need 6 daily answers per linked prompt and engine after a change, on at least 4 separate days spanning at least 10 days. Each side counts one answer per prompt, engine and market per day.
A change is reported only if it is at least 15 points, passes a two-sided Fisher exact test at p < 0.05, and similar unchanged prompts didn't move the same way. Results are observations, not proof that the change caused them.
Compare names a winner only when both sides have at least 10 answers and the same exact test separates them.

AI reasoning by task

Each AI step uses the reasoning effort its job needs. Checks that run on every collected answer use low effort. Decisions about competitors, recommendations and localisation use medium. Anything that tells a client something is wrong, or writes client-facing copy, uses medium or high.

Task	Effort	Why
Brand and competitor names in an answer	low	Literal extraction of names from one answer into a schema.
Selection signals in an answer	low	Structured signals from one answer; runs on every collected answer.
Diagnostic score	low	Rubric scoring of one answer against fixed criteria.
Sentiment	low	Tone classification of the sentences that name the brand.
Recommendation evidence	low	Literal evidence extraction; a deterministic fallback covers failures.
Track: business-driver associations	medium	Brand–attribute associations with polarity; negation and comparisons decide whether an association is favorable.
Competitor verification	medium	Decides which named companies are real rivals; drives share of voice and competitor-owned sources.
Accuracy: claim extraction	medium	Splits an answer into checkable claims about the brand.
Accuracy: verdict against confirmed facts	high	Tells a client an AI answer is wrong about them; a false error is costly.
Prompt labelling	low	Labels a prompt's intent and stage from a short text.
Brand profile from the website	medium	Synthesises a brand profile from its website.
Fact base draft	medium	Drafts verifiable facts from the brand's pages for human review.
Arabic aliases	medium	Transliterations and spellings need linguistic judgement.
Arabic prompt variants	medium	Localised prompts must keep the buyer's intent and dialect.
Recommendation verification	medium	Second-pass check that each action follows from the evidence.
Writer: outline	high	Structure decides whether the page answers the buyer's question.
Writer: draft	high	Client-facing copy that must stay inside verified facts.
Writer: chat	medium	Interactive; answers must stay grounded without long waits.
Writer: claim audit	high	Audits every factual claim in a draft before approval.
Writer: evidence check	medium	Compares a claim with the text of one cited page.
Writer: quality review	medium	Quality and business-fit review of a draft.

What we don't claim

We record mentions and links in answers. We don't measure clicks, traffic or revenue, and a visibility change isn't a revenue change.
Content-characteristic scores are our analysis, not the engines' ranking factors or proof of why a source was chosen.
In-country collection is verified where the provider reports the egress country; unverified runs are labelled.

Questions about the method? Contact us.