Sequence the Work, Set the Clocks

The executive question that kills AI SEO programs is “why aren’t we cited yet?” — asked in month two, about work that physically cannot show results before month six. Sequencing is therefore two designs in one: the order you build capabilities, and the expectations you set for when each one pays.

The two clocks

AI visibility moves on two layers with different physics.

The retrieval layer is what an assistant can look up at answer time: a search index, a live page fetch, a product feed. It refreshes on crawler cadence — hours to days — and push mechanisms like IndexNow can notify the Bing index (which grounds both Copilot and ChatGPT search) at publish time. Work on this layer can show movement within weeks. The distinction is unpacked in LLM Training vs Real-Time Indexing.

The model layer is what the underlying model believes from training data. It moves only when a model is retrained or replaced. No page edit reaches it; only sustained entity signals and earned coverage migrate into the next training run. Brand-perception work is a multi-quarter bet by construction, and any plan that promises otherwise is misdescribing the system it claims to optimize.

Sequence

First: measurement baseline. Before optimizing anything, stand up the scoreboard — prompt corpus, per-surface baseline, first-party data (Bing Webmaster Tools’ AI Performance report is free and platform-native). Optimization before baseline produces a program that can never prove it did anything. The measurement guide covers the design; do it first anyway.

Second: the technical floor. Verify crawler access against server logs — declared policy and actual bot behavior diverge more often than anyone admits, usually via a WAF rule nobody remembers. Confirm indexability on both Google and Bing. Validate structured data and fix schema-versus-page contradictions. This is unglamorous, fast, and inherited by every surface at once. The infrastructure checklist is the verification pass.

Third: content and answer structure. Restructure the pages that answer your buyers’ highest-value questions; retrieval-layer surfaces reward it on recrawl. This work compounds but starts paying in the first quarter.

Fourth: entity and earned media. Slowest, most durable, and the only lever that reaches the model layer. Start it early precisely because it pays late.

International: sequence by market tier, not all at once. AI-assistant adoption is uneven across markets; identical global rollout overspends where adoption lags and starves the markets pulling ahead. Tier markets, set tier-appropriate goals, revisit semi-annually.

Setting expectations that survive contact

Three commitments to extract from leadership before the first invoice:

• Rolling windows, not snapshots. AI answers are non-deterministic and cited sources churn. Report 90-day rolling movement, never single-month readings.
• Per-surface timelines. Retrieval-layer surfaces move in weeks; systematic programs on the Bing/Copilot stack are typically a one-quarter feedback loop; model-layer perception is six months to a year. One blended “AI visibility” date is fiction.
• Both reasoning modes count. Semrush found ChatGPT’s two reasoning modes cite mostly different domains — a brand can win the mode it measures and stay invisible in the one its buyers use. Sampling breadth is part of the timeline conversation, not a refinement for later.

The shape of the first two quarters

A realistic composite: quarter one is baseline plus technical floor plus the first content restructures — expect measurable retrieval-layer movement on Perplexity and Bing-grounded surfaces late in the quarter. Quarter two is content at scale plus entity work in flight — expect citation share to move on the surfaces you scoped, and expect leadership to ask about the model layer, which has not moved yet. If the QBR deck does not pre-empt that question with the two-clock model, this page has failed you.