How it works
Method, not black box.
Every score traces back to a page, a tag or a model answer. Every run starts with an estimate you can refuse. This page shows the mechanics.
Three ways in
Start from a sector question set, your own prompts, or a product feed.
System Prompts
Measure the governed prompt set saved for your sector and brand: Turkish banking ships 168 prompts across retail, SME and corporate segments.
Custom Prompts
Ask ad-hoc questions without touching the saved taxonomy. The right way to test a campaign or a product launch.
Data Feeding
Audit a product feed (CSV, JSON, XML) with no crawling at all, for sites that block crawlers or catalogs that have not shipped yet.
One operating loop, six stages
Every analysis moves through the same durable pipeline.
The stage is derived server-side from durable records, so it survives restarts, devices and sessions. At each stage you decide; AgentLens records.
Stage 1 of 6
Configure
Brand, prompts, rival pool, AI surfaces and cadence live in Brand Management. Saving creates a new prompt-set version; history is never rewritten.
- You decide
- Which questions matter, which rivals count, which surfaces to measure.
- Output
- A versioned benchmark definition.
Stage 2 of 6
Measure
A spend-free estimate shows pinned models, expected calls and a hard cap. Only after you confirm does a run ask the real vendor models.
- You decide
- Whether the estimate is worth confirming.
- Output
- Answers per prompt, per surface, with provenance recorded.
Stage 3 of 6
Diagnose
Four-pillar score with confidence intervals, an evidence map from every score back to its page or answer, prioritised issues and guides.
- You decide
- Which root causes to act on first.
- Output
- A ranked action queue built from what was actually recorded.
Stage 4 of 6
Improve
Fix Engine and Content Optimizer generate deployable artifacts from the run: JSON-LD, llms.txt, robots.txt, meta tags, rewritten content, feeds.
- You decide
- Which artifacts to take.
- Output
- Preview, copy, download. Nothing is pushed to your site.
Stage 5 of 6
Approve & publish
You review; your own pipeline deploys. The treatment is recorded as an intervention so the next measurement can be compared against it.
- You decide
- What ships, and when.
- Output
- A dated intervention record.
Stage 6 of 6
Re-measure
Scheduled cycles report what changed and whether it beat sampling noise. Alerts reach high severity only when a significance test rejects noise.
- You decide
- The cadence and the delivery channel.
- Output
- A movement you can defend in a meeting.
One scorecard, four pillars
How each pillar is measured.
Transparent formulas. Each pillar is scored 0–100 and carries equal 25% weight in the overall readiness score.
Overall readiness score
Four pillars, equal weight, each 0–100. Every score ships with a 95% confidence interval and its sample size.
(Schema × 0.25) + (Content × 0.25) + (Share of Model × 0.25) + (Shopping × 0.25)
- 70–100
- Excellent · Well optimised for AI discovery
- 40–69
- Moderate · Room for improvement
- 0–39
- Poor · Significant optimisation needed
Variance control
- Multi-trial averaging (1–5 trials)
- Pinned model versions
- Deterministic sampling
- 95% confidence interval on every metric
- Pillar 125%
Schema Coverage
JSON-LD structured data: presence, correctness, completeness.
(pages with valid JSON-LD ÷ pages crawled) × 100
Per-page validation of Product, Organization, BreadcrumbList, FAQPage, Offer and Review, with issue-level diagnostics.
- Pillar 225%
Content Structure
Semantic HTML, meta quality, heading hierarchy, llms.txt.
semantic HTML (25) + structure (25) + meta (25) + llms.txt (25)
HTML5 landmarks, H1→H2→H3 hierarchy, title, description, alt and OG tags, and a compliant /llms.txt at the root.
- Pillar 325%
Agentic Brand Familiarness
How often AI mentions your brand organically: Share of Model.
(mentions found ÷ prompts tested) × 100
Runs on OpenAI by default; benchmark runs add Claude, Gemini, Grok and Google AI Mode and record whether each answer browsed.
- Pillar 425%
Shopping Assistant
Position and frequency in AI product-recommendation shortlists.
Σ (points per query × position multiplier)
Rank 1 = 100%, rank 5 = 60%, rank 10+ = 0%. Includes product grounding and attribute-parity checks.
You see the bill before the run
Estimate, confirm, run.
Tenant-scoped SaaS with explicit spend control. Your keys, your data boundary, your export formats.
Spend-free estimate
Pinned models, expected calls, estimated cost and a hard cap at 115%. No run, no reservation, no provider call yet.
Smallest detectable change
The estimate tells you the least week-over-week movement this run could confirm, before you spend anything.
Explicit confirmation
A run is created only with a current, single-use estimate and your cost confirmation. A changed shape, an expired estimate or a cap above 115% is rejected server-side.
Measurement probes always run on the real vendor model, or on Google AI Mode as a live search surface. Analysis-side drafting and classification can run on the platform model plane; measurement never does.
What a measurement tells you
Being mentioned is not being recommended.
Four judgements, each versioned and evidence-preserving, so a number on the dashboard can be traced to the sentence that produced it.
Mention quality
mention-quality-v1Role, rank (1–100) and sentiment for every judged answer. Evidence must be a verbatim quote that names the brand; otherwise the judgement is discarded, not guessed.
Product comparison quality
comparison-quality-v1When a model compares you against a rival: who was preferred, which factors drove it, and whether the cited evidence supports the outcome. Target preferred · competitor preferred · conditional · mixed · no clear winner.
Surface provenance
per answerEvery answer records where it came from: model memory with no browsing, or a live-search surface such as Google AI Mode. Grounded and ungrounded answers are shown as a split, never blended into one average.
Rivals observed
from evidenceRivals the models compared you against, but you are not tracking yet, surface with their evidence. The rival pool grows from observation, not guesswork.
Start with one analysis
See the method run on your own domain.
One confirmed run returns a four-pillar score with its evidence, mention quality per answer and a prioritised action queue. We reply within five business days.