Measurement first.
Everything else follows.
The industry around AI visibility runs on anecdotes — a screenshot of one good answer, a case study with no control. Citehelm's methodology exists to replace hope with a testable transfer function: which actions, on which pages and sources, change which answers.
Sentinel sets
Everything starts with the questions. We map the prompts your buyers actually use — by language, intent and funnel stage — into clusters, then designate sentinel sets: representative prompt samples that are re-run on a fixed cadence. Weekly sentinel runs catch movement early; a monthly full-protocol run covers the entire taxonomy across all monitored engines.
Prompt clusters are split into test and control groups. Interventions target test clusters only. If test moves and control doesn't, the intervention worked; if both move, the engine changed under everyone's feet — and we tell you that instead of taking credit.
Citation-set analysis
For every cluster we record which sources the engines retrieve and cite, and how stable that consensus is. Clusters divide into open (no entrenched citation set — fastest to win) and locked (a stable consensus dominated by strong sources — a longer, different game). Pricing and expectations are calibrated to that split, established during the baseline audit.
We also separate the two mechanisms that get conflated everywhere else: Play A, own-domain retrieval and citation, and Play B, being named by third-party sources the engines already trust. They respond to different interventions and are tracked as separate workstreams.
A versioned action log
Every fix is a logged, dated action tied to the clusters it targets: retrievable content built to answer, technical eligibility work (crawlability, structure, hreflang), verifiable authorship and entity infrastructure, and earned presence on the citation sources themselves. No invisible "optimisation" — you can read the log.
90-day windows, honest deltas
Interventions are evaluated against their measurement window with the test/control split intact. Reports show citation share per cluster before and after, per engine — including the null results. Silent decay is monitored the same way: a citation you hold today can erode, and the sentinel cadence exists to catch it before your pipeline does.
Direct questions, direct answers
- Is this just SEO with a new name?
- It overlaps with SEO the way aviation overlaps with sailing: shared physics, different craft. Classic SEO optimises for ranked lists of links; we optimise for inclusion in the retrieved-and-cited set behind a generated answer. Technical hygiene still matters, but the diagnostics, the interventions and above all the measurement are different.
- Can you guarantee that AI engines will cite us?
- No — and no honest operator can. Generated answers are stochastic: the same prompt can produce different citations on different days. What we guarantee is the discipline: a real baseline, interventions logged as testable actions, and verification against control prompts, so you always know what moved and what did not.
- Why a retainer instead of a one-off audit?
- Because citations decay silently. An audit describes a moving system once; by the time you act on it, the answer set has shifted. Monitoring plus remediation is the only structure that matches how AI answers actually behave over time.
- Which engines do you monitor?
- The set evolves with usage share. Currently: ChatGPT (including its search mode), Perplexity, and Google’s AI answers, with others added when they matter in your market and language.
- How long until results are visible?
- We work in 90-day measurement windows. Open prompt clusters — where no strong citation consensus exists yet — can move within one window. Locked clusters, dominated by entrenched sources, take longer and we tell you which is which after the baseline.