Frequently asked

Questions we hear (and answer) a lot.

Short. Opinionated. Where the long answer matters, the link goes back to the source.

What is Combot

What problem does Combot solve?

The BI stack from the Google era doesn't tell you whether your brand is recommended by ChatGPT, Claude, Gemini, or Perplexity for a category prompt that drives revenue. Combot does. It runs a version-controlled prompt portfolio nightly against the major frontier models, decomposes the answers by funnel layer and knowledge mode, and surfaces one number — AI Share of Recommendation — with the drill-down to act on.

How is Combot different from Profound, Otterly, Goodie, Athena?

Those tools measure mentions, citations, and share of voice. Combot measures weighted recommendation — whether the model said "I'd choose this brand because of X" rather than "this brand exists." Different signal, different fix family. Combot also decomposes each outcome by knowledge mode (Memory, Search, Fetch) so the team knows whether the fix is a Wikipedia edit, a content rewrite, or a deploy that broke SSR.

Who is Combot for?

Brands with >€10M revenue that have a measurable category-discovery surface in AI answers. The current pilot serves an ecommerce brand; the platform extends to B2B SaaS, travel, and any consideration-heavy category. Smaller brands without a portfolio of category prompts can still use Combot, but the ROI sharpens above a threshold of AI-driven discovery traffic.

Is Combot a SaaS or a service?

Both. The dashboard is SaaS. The first 90 days of any new client engagement is co-piloted by the Combot team to calibrate the prompt portfolio, channel weights, and finding pipeline. After calibration, the dashboard runs itself; the team reviews outliers.

AI visibility basics

Is "AI visibility" just SEO with extra steps?

No. Classic SEO optimises for one mechanism: a search engine indexes your pages and ranks them in response to a query. AI visibility optimises for three orthogonal mechanisms — what the model remembers from training, what it retrieves via web search, and what it can fetch from a specific URL. Each fails differently. Each rewards different work. The teams, tools, and time-scales are different.

Why 7 layers?

"Are we visible in AI?" is seven different questions, not one. A brand can be known to the model but never recommended; recommended but with wrong facts; recommended with right facts but unreachable for the action. The 7-layer funnel separates the failure modes so each gets its own fix.

What is the north-star metric?

AI Share of Recommendation — the percentage of relevant prompts where the brand is recommended as a viable or preferred choice, weighted by prompt value, model importance, geography, and funnel intent. In v0.2 it also decomposes by knowledge mode (Memory-SOR, Search-SOR, Fetch-SOR) so the headline number is explainable.

Do I have to learn 13 things to use Combot?

No. The dashboard surfaces the one number first. The 7 layers are how the dashboard diagnoses; you don't need to memorise them to act on a finding.

How often does Combot probe?

Uniformly weekly. We considered tiering by prompt-length (cheap weekly short + expensive monthly long) but rejected it. Short-answer constraints act as a truncation filter — models drop secondary brands to save tokens, producing false-negative trends on the headline metric. We run the same declarative prompts every week and trust the consistent signal.

What's a "declarative" prompt? Why not just ask the model openly?

Open-ended prompts give variable-length answers with volatile mention counts. A model might cite a brand 4 times in 500 words and 0 times in 50 words on the same question. We use declarative constraint patterns — for example, "Name your top 3 in this category. One sentence per choice." This stabilises the metric by turning continuous mention counts into binary inclusion, without distorting what the model would naturally pick.

Knowledge Modes

What's a "knowledge mode"?

A mode is one of the three independent retrieval mechanisms a model can use to answer a question about your brand: Memory (its training-time knowledge), Search (a live web search via its built-in tool), or Fetch (direct read of a specific URL). The mode determines which optimisation lever applies. See the full post.

How do I optimise for Memory mode?

Slow, compounding work: Wikipedia + Wikidata entity discipline, schema.org Organization JSON-LD, sustained coverage in publications that become pretraining data (HN, TechCrunch, FT, Bloomberg), consistent NAP across the open knowledge graph. Time-to-impact: months to years (gated by vendor model-release cycles).

How do I optimise for Search mode?

Medium-speed, content-shaped: classic SEO foundations (rank well first), /llms.txt, AI-bot allowlist in robots.txt, answer-first 40-60 word summaries near the top of each page, structured comparison tables. Time-to-impact: 1-6 weeks.

How do I optimise for Fetch mode?

Fast, engineering-shaped: server-side render (most AI bots don't execute JavaScript), TTFB under 800ms, semantic HTML (<main>, <article>), JSON-LD with the facts the model needs, no cookie wall blocking text, no PDF-only product specs. Time-to-impact: 1 day to 2 weeks. Biggest quick wins for technically-broken sites.

Are the modes equally weighted?

No. The weighting depends on intent. Reputation prompts weight Memory higher. Category-discovery prompts weight Search higher. Product, pricing, and action prompts weight Fetch higher. Combot ships defaults; clients can override per prompt family.

The pilot

What's the pilot setup?

8 weeks. One pilot brand at a time. Combot provisions a dedicated GCP project, builds the prompt portfolio from your GSC + sales-call + Ads-search-term data, runs the nightly pipeline, and delivers a weekly Pulse review. After week 4, the channel weights and finding pipeline are calibrated against real data.

What does the pilot cost?

Talk to [email protected]. Price depends on prompt portfolio size, model coverage (3 models default vs adding OpenAI / Perplexity / xAI), and the depth of Monitoring (top 50 URLs vs full site). For an Elisa-sized ecommerce site, expect cloud + LLM costs to be the floor.

How long until I see signal?

Week 1: prompt portfolio + baseline AI SOR. Week 2: first nightly anomaly findings. Week 4: calibrated channel weights + finding-to-fix pipeline. Week 8: dashboard runs itself with weekly human review.

What do I need to bring?

Read access to your GA4 (BigQuery export), Search Console (Bulk export), and ideally Google Ads + Merchant Center. Write access only if you want findings to flow into Linear / Slack / Airtable. A 30-minute calibration call to align on prompt portfolio + channel weights.

Will I see daily fluctuations?

Weekly aggregates only. Day-to-day LLM answers can vary 30-60% from sampling temperature alone — that's industry-standard noise, not signal. We aggregate weekly to smooth the variance, surface the trend, and alert on >2σ moves. If you need finer granularity than weekly, talk to your Combot contact — daily tracking is available but inherently noisier.

Technical setup

Where does the data live?

Each client gets its own GCP project. Datasets are clustered by client + entity (BigQuery does the work). No cross-tenant joins. Audit trail per query. R2 holds large blobs (probe transcripts, URL snapshots, screenshots).

How is Combot Chat verified?

Every analytical prompt fans out to two independent models in parallel — Claude Sonnet 4.6 and Gemini 3.1 Flash-Lite. Both have access to the same toolset (vendor web_search + web_fetch + our BigQuery + URL history). A third model — Claude Opus 4.7 — synthesises the two transcripts under a strict prompt that surfaces agreements + disagreements explicitly. Hallucinations are caught by cross-vendor heterogeneity; averaging is forbidden. The user gets one final answer with the option to expand both runner transcripts.

Why two models, not three?

Heterogeneous-vendor pairs catch the vast majority of single-vendor hallucinations. Adding a third runner increases cost ~50% with diminishing returns on divergence. We may revisit if the team finds two-runner outputs collude.

How fast is Combot Chat?

Verify mode: 12s p50 / 20s p95 end-to-end. Express mode (Opus solo with tools): 5s p50 / 12s p95. Both budgets are targets to validate against production data; latency settles after the first 100 calls per client.

What's the EXPRESS mode toggle?

A switch in the Combot Chat panel header. Verify (default) runs the full dual-runner + synthesiser architecture: ~12s, ~$0.30/prompt, surfaces disagreements. Express skips the parallel fan-out and uses Opus 4.7 alone with the same toolset: ~5s, ~$0.15/prompt. Use Express for fast iteration and low-stakes lookups; Verify for board-grade analytical answers.

Alerts & notifications

When will I get a browser push notification?

Only for CRIT (critical) severity alerts. Medium and high anomalies stay quietly in the in-app inbox to avoid alert fatigue. The browser permission prompt only fires when you explicitly click Enable on the in-app banner at the top of /alerts — never on login.

What's the difference between Acknowledge, Snooze, and Resolve?

Acknowledge: "I've seen this and I'm investigating." Alert stays open in the inbox but stops sending notifications. Snooze: "Remind me later." Suppresses the alert for 4h / 24h / 7d / 30d. Resolve: terminal state — "this is fixed." Audit-logged and removed from the open list. If the same trigger fires again it creates a new alert.

How does Combot detect anomalies in my AI visibility?

2σ deviation from a 7-day rolling baseline per (URL × bot × model × knowledge mode). If your Sonnet-Search-mode SOR for a category-discovery prompt drops 2 standard deviations below the rolling mean, that's an anomaly. Thresholds are configurable per intent family; defaults are tuned conservatively to avoid noise.

Vendor & model

Which AI models does Combot probe?

Default team: Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro. Extendable to Perplexity Sonar, xAI Grok, and any new frontier model with public API access. Per-model SOR + per-(model, mode) cells live in the cube.

Does Combot send my data to OpenAI / Anthropic / Google?

Yes for tool prompts and synthesis; no for any data stored in BigQuery beyond the specific tool call. Each model's API is invoked over HTTPS with no training-data opt-in. Per-vendor data handling: Anthropic + Google do not train on API traffic by default; OpenAI's API also opts out of training by default.

What's the relationship between Combot and combot.dev?

combot.ai is the public face of the product. combot.dev is the engineering domain — APIs, downloads, internal dashboards. Clients land on {client}.combot.ai for their dashboard.

Why Anthropic + Google for the dual runner, not OpenAI?

Vendor diversity in the runner pair matters more than picking the "best" model. Anthropic's Sonnet and Google's Gemini have completely different retrieval backends (Brave vs Google) and architectural lineage, which maximises catching single-vendor hallucinations. OpenAI sits in the model team that gets probed against, not in the runner pair.

Don't see your question? Email [email protected] — answers usually land in the next FAQ revision.