On the board
The 40 that score highest become the targets. The rest stay tracked, so a bad answer shows up before a lost deal does.
Average visibility out of 100 across seven engines, 18 to 24 September 2026. Visibility is our tracking tool's score for how often and how prominently an engine names a brand.
A small set can't tell a real change from the engines' own noise.
Re-checks of an unchanged brand that moved 2 points or more. Salienz, Audit Size Test, October 2026.
Never the bare name. "Sequence" alone returns genetics and mathematics, so every branded question carries a qualifier.
On the board
On the board
Fixed for the term
On the board
On the board
Choosing a product, or still learning about the problem?
How much proof that buyers ask it, from the interview, search data, threads and reviews?
Can you credibly be named in this answer within the term?
Near-duplicates are merged, so two targets are never spent on one answer.
Questions that touch a weak spot in the product stay tracked, never targeted.
Where engines answer with sources rather than vendors, a citation of your site is what counts.
On the board
Stripe's score out of 100 on the same questions, Sequence's first week of tracking.
Answers citing a source
Security awareness answers naming the leader
Salienz, AI Security Shortlist Study, October 2026.
On the board
Comparison pages, alternatives pages and answer pages, each opening with a direct answer, dated and sourced so an engine can lift a passage whole. Pages that rank but are never quoted get rewritten.
10 cite-ready assets a month on Domination, 15 on Expansion, plus up to 4 or 8 rewrites of pages you already have.
The roundups, listicles and threads the engines already cite, picked from the baseline. In one six-day baseline, a single 2026 roundup was cited 126 times.
Up to 15 placements across the term on Domination, 25 on Expansion, with guest posts and Reddit under real names.
Technical fixes on a fixed cycle, because an engine can't quote a page it can't crawl or render. Entity work alongside it: what the engines get wrong about you, corrected at the source they took it from.
2 technical fix cycles a quarter, and up to 90 internal links a quarter on Domination, 180 on Expansion.
The same method runs at 100 questions on Domination and 200 on Expansion.
DominationExpansionOn the board
Questions set out to be won by day 90. A target, not a guarantee: engines re-rank without notice.
Salienz builds the prompt set from a client interview, the client's own data, and outside research in DataForSEO, Google autocomplete, Reddit, social platforms, review sites, competitors' pages and live engine answers. The default set is 90 unbranded and 10 branded questions per hundred, each with its evidence attached, agreed with the client before it runs.
Unbranded questions are where buyers build shortlists before they know a vendor. In September 2026, Sequence averaged 83.7 out of 100 on the ten tracked questions naming it and 0.9 on the ninety that did not, so a set weighted toward branded questions would have reported success while missing the problem.
Each tracked question is scored out of 100: 40 points for closeness to a purchase, 35 for evidence of demand and 25 for whether it can be won in 90 days. The top 40 of 100, or the top 60 of 200, become targets. Near-duplicates are merged, and questions touching a known product weakness are tracked but not targeted.
Salienz tracks five engines: ChatGPT, Claude, Perplexity, Microsoft Copilot and Google. Google's three surfaces, AI Overviews, AI Mode and Gemini, are scored separately because they answer differently, so reports show seven. In one September 2026 week, Stripe scored 19.1, 23.3 and 30.0 on those three Google surfaces for the same questions.
The Domination plan sets out 30 to 40 questions won by day 90 against 40 targets, and Expansion sets out 45 to 60 against 60, measured on the same fixed set every month. Neither is guaranteed, because engines re-rank without notice. The scope, the baseline and a monthly report that shows a flat month as flat are.
Run the method on your category first.
The Foundation audit: 100 questions across five engines, $800, delivered in seven business days.