The engines behind AI answers
AITWIRE polls 9 AI systems covering 90%+ of AI answers — the assistants and answer engines your customers actually use, such as ChatGPT, Perplexity, Claude, Gemini, and Grok. Google AI Overviews is additionally tracked through blended Search Console impressions.
Training-prior answers versus live-search answers
Not all AI answers are produced the same way, and AITWIRE distinguishes them:
- Training-prior engines answer from what the model learned during training. They change only when the model is retrained, so correcting them is a longer game. As of early 2026 the standard Grok chat mode is one of these: xAI retired its live-search auto-grounding, so bare Grok now answers from its training prior.
- Live-search engines retrieve the web at answer time and can cite their sources, so they reflect your freshly published facts within days of a re-crawl. Perplexity and ChatGPT Search work this way — and AITWIRE now also polls a live-search variant of Grok (
grok_search) that uses a live web-search tool and returns citations, tracked as a separate, opt-in line on top of the core set so it never muddies the training-prior Grok trend.
Knowing which kind of engine gave an answer tells you what to expect from a fix: live-search engines move within days of a re-crawl; training-prior engines move on the model retraining schedule.
How probes are allocated
By design, AITWIRE concentrates your probe budget where it carries the most signal rather than spreading it thin:
- A fixed sentinel core of your highest-priority questions — your core brand-identity and pricing facts — is measured every cycle, so your headline trend line never loses continuity.
- Event-driven probes keep full density: anything steered by an active campaign, and any question inside a verification window after you publish a fix.
- The remaining questions rotate on a deterministic schedule, so the whole set is covered over successive cycles without over-sampling any one area.
This is a measurement-design choice, not cost-cutting: steady-state accuracy drifts only slowly, so anchoring the questions that matter and rotating the rest measures what counts without adding noise.