The situation
Intryc sells AI quality assurance to customer support teams. Its buyers, VPs and directors of CX, increasingly start vendor research by asking an AI which QA platform to buy. The sentence the AI writes back is the new shortlist.
Intryc was already in that conversation, and AI engines already trusted its content. But that trust wasn't turning into recommendations often enough, and each engine told a different story about where Intryc stood.
A pre-engagement audit put the full picture on one page, and the conclusion was clear. The content existed. It just wasn't structured for machines to retrieve it.
What was in the way
None of this was a content problem. It was the drift every fast-moving product accumulates, and it stays invisible until you measure AI answers directly.
- Citations pointing at retired pages
As the product evolved, pages moved. AI kept citing the old URLs, so hundreds of AI-referred buyers landed on pages that no longer existed.
- Credit going to the wrong name
AI was citing Intryc's own content in answers that recommended someone else.
- Pages written for people, not machines
Pages that read well to a buyer were hard for a model to parse, quote and attribute.
Every one of those dead citations was a buyer who asked AI for a recommendation, clicked, and hit nothing.
The strategy: four decisions
- Treat each AI engine as a separate market
Each engine builds its answers from different sources, so each needed a different fix. One blended "AI visibility strategy" would have wasted half the budget.
- Defend existing citations before chasing new ones
Citations aimed at dead pages, plus more on pages AI could barely parse: hundreds of citations of leverage available without writing anything new. Net-new content is the expensive move. It came second.
- Attack where a competitor is present and the client is absent
Not keyword volume or topic interest. The backlog was every tracked buying prompt where AI named a competitor and not Intryc. That is the whole addressable gap, and it is finite and countable.
- Take the contested ground directly
Where AI recommended a competitor instead, more thought leadership wasn't the fix. Intryc got a dedicated, honest page built to win that exact answer.
How we did it
- Measure what had never been measuredWeek 1
A tracked query bank mapped to the real buying journey, an AI visibility scorecard, and a continuous benchmark sampling 1,100+ AI answers per run across ChatGPT, Claude, Gemini and Perplexity, with head-to-head tracking against the competitors buyers actually compare.
The output was a frozen baseline. Nothing shipped until the starting point was on record. Without it, every later claim is a story.
- Fix the plumbing before writing a wordWeeks 1–2
Dead citations redirected to live pages. Attribution fixed so AI credits the right vendor. Structure and markup added to the pages AI was already reading.
Nobody sells this part because it is unglamorous. It is also the part with the shortest distance between the fix and measured movement.
- Build the claim spineWeek 2
One source of truth for positioning, voice and verified customer proof, plus a winning narrative for every contested buying question: what AI says today, what is true, and the sentence that should be cited instead.
Every page draws from it, so every page says the same thing the same way. That consistency is what AI rewards.
- Produce against ranked gaps, not a content calendarWeeks 2–4
Every play is ranked by expected impact before it enters the queue. Wave 1 shipped 14 plays: new comparison and category pages, upgrades to pages AI already cited, and structural fixes.
Every page clears quality gates for voice, structure and accuracy before it goes live, including review by a second, independent model. Those gates have already caught errors a human skim would have missed.
- CompoundContinuous
Every correction becomes a permanent rule. The system that wrote page 30 is measurably better than the one that wrote page 1, because it can't repeat a mistake it has already made.
The results
Measured on the KnitKnot AI Presence benchmark: 1,144 sampled AI answers at baseline and 1,129 on day 28, across ChatGPT, Claude, Gemini and Perplexity.
Visibility counts non-branded prompts only. Engine scores are per-engine averages on the same benchmark.
Core visibility metrics
| Metric | Baseline | Day 28 | Move |
|---|---|---|---|
| AI Presence Score | 39.7 | 48.3 | +8.6 pts (+22%) |
| Visibility (non-branded) | 17.98% | 37.13% | +19.2pp (2.1×) |
| Share of voice | 12.31% | 14.88% | +2.6pp (+21%) |
| Sentiment when named | 96.5 | 97.0 | Holds |
| Average rank when listed | 2.0 | 2.5 | −0.5 |
| Answers sampled | 1,144 | 1,129 | 4 engines |
Sentiment was already 97 out of 100. AI liked Intryc fine. It just wasn't mentioning them. Average rank slipped because Intryc started appearing in answers it had never been in before. Being named fifth in an answer you used to be missing from is a gain.
Per-engine average score
| Engine | Baseline | Latest | Move |
|---|---|---|---|
| Gemini | 44 | 60 | +16 |
| Perplexity | 32 | 43 | +11 |
| Claude | 47 | 51 | +4 |
| ChatGPT | 36 | 40 | +4 |
The baseline flagged Perplexity as the weakest surface with the most upside. It delivered the second-largest gain.
What got installed: the scoreboard is the deliverable
Query bank, scorecard, a 4-engine benchmark that re-runs on its own, and head-to-head tracking against key competitors.
Dead-citation recovery, attribution fixes, schema, llms.txt and retrieval-chunk hygiene. Citations already earned, working again.
Brand brain, winning narrative, voice guide and a verified proof bank. Every future page draws on one source of truth.
Ranked gap plays, quality gates and cross-model review. Pages built for the prompts where competitors win.
A lessons store, weekly strategy reflection and decay monitoring. The client approves the claim bank, merges the PRs and makes the calls only a founder can make: a few hours a month. The system does the rest.
