What this product is for

llmeknow measures AI Perception: how large language models describe, compare, recommend, or omit brands, organisations, people, and ideas when people ask everyday questions. You set up a market and a campaign, run a wave across chosen models and audience segments, then read the responses, entities, sources, and reports those waves produce.

Signing up needs a card, verified with a R1 charge that's credited straight back to your wallet, then you have 30 days to explore. Running a wave still needs wallet credits, because each response is a real model call.

The core loop

Market: the category and geography you care about (for example car insurance in South Africa, or municipal housing policy in a city).
Campaign: the questions, entity classes, segments, models, and intent for that market.
Wave: one full execution of the campaign. Repeat waves over time when you want trend, not a one-off snapshot.

Home walks you through this when the workspace is empty. Glossary in the sidebar defines the same terms in one place.

Choose a path by objective

You can use the same machinery for different jobs. Start from the outcome you need, then set campaign intent, questions, and analysis tools to match.

Brand and marketing work

Use this when you care how models talk about your brand versus competitors: who gets recommended, which attributes stick, and where you are missing from the answer set.

Set campaign intent to Representing your brand or Tracking competitors.
Keep questions neutral enough that you are not priming one brand; see avoiding context leak.
Lean on entity classes for brands and attributes, then clean entities before trusting Share of Voice.
Use segments when the interesting gap is by audience (for example first-time buyers versus switchers), not when every demographic dial is turned on at once; see designing segments.

PR and reputation work

Use this when the question is how models frame your organisation after news, crises, or campaigns: tone, associations, and which stories get repeated.

Prefer Neutral research or a short custom intent that names the reputation question you are answering.
Write questions the way a journalist or concerned citizen would ask them, not as brand slogans.
Watch sources and source quality as carefully as the answer text; reputation often rides on what the model cites.
Use co-occurrence and relation views to see what names travel with yours, and keep co-occurrence distinct from direct attribution.

Socio-political and public-affairs work

Use this when the subject is policy, institutions, public figures, or contested social questions rather than product brands.

Markets still need a clear geography and category (who is being asked about, and in which place).
Questions should stay open and parallel across actors; naming one party and not others tilts the wave the same way brand context leak does.
Entity classes often expand beyond brands into organisations, people, programmes, and issue labels. Treat cleanup as part of the method, not an afterthought.
Narrative tracking is usually the right extra axis here: you define statements you care about (for example a policy claim), then score how strongly responses agree or disagree with those statements across models and segments. Attach a taxonomy when you need stance over time, not only entity mention share.

Parts of the product you will use most

Segmentation

Segments are audience groups defined by attributes (demographics, attitudes, context). The platform generates synthetic individuals inside each segment and asks your questions as those people. Use few attributes that match the decision you need; overloaded personas overfit. Always keep a baseline (no segment) so you can see what the models say without audience context.

Entity classes and cleanup

After a wave, extraction pulls named things into classes (brands, attributes, and any custom classes you configured). Analytics join on those classes. Before you quote Share of Voice or rankings, merge duplicates, drop noise, and reclassify mis-typed entities. When you change the entity set, refresh the report after cleanup so charts and write-ups match the cleaned data.

Narrative tracking

Narrative tracking scores stance toward a curated set of statements (theme → topic → statement). It sits beside entity tracking when the question is agreement or disagreement with claims, not only whether a name appeared. Use it for policy debates, reputation claims, and message testing. Attach a taxonomy to the campaign, optionally backfill older responses, then read stance by model and segment on the Narratives views.

Questions, repetitions, and models

Question design drives almost everything downstream. Pair intent with careful phrasing (campaign intent, context leak). Use enough repetitions that a single lucky or unlucky answer does not define the result. Pick models that matter to your audience; compare them rather than assuming one voice equals the market.

Sources and reports

Models that search the web leave sources on responses. Treat those as evidence to inspect, not as proof by volume. The AI overview and exportable reports summarise a wave for stakeholders; they are only as good as the cleanup and intent behind them.

A practical first week

Create one market with a real geography and category you already understand.
Create one campaign with a clear intent and a short question set (five strong questions beat fifteen vague ones).
Add one or two segments only if the decision depends on audience difference; otherwise run baseline alone first.
Run a wave when you have wallet credits, then clean entities before sharing numbers.
Add narrative tracking if your questions are about claims and stance, not only brand presence.
Come back for a second wave when something in the world changes (launch, crisis, policy shift), so you can read movement rather than a single cut.

Getting started with llmeknow