Every tactic in this book pulls from the same budget, and strategy is deciding where that budget goes. Two inputs should drive the decision: how AI answers behave in your category, and where you currently stand in them. Both are measurable before you spend anything, and this chapter gives the scoring logic, the four starting positions, and the sequencing that follows from them.
Score the category before scoring yourself
Categories do not behave alike in AI answers. In OppAlerts' study across 100 industries, the top five brands in a category held 49% of all AI answer share, against 24% for the top five in Google organic in the same industries.The OppAlerts ranking-factors study: 403,000 prompts, 100 industries, 13 models, 1,100 personas. Method and per-industry benchmarks are in The Research Behind This Guide, part of the AI Search Visibility research. That 49% is an average, and the industries around it spread widely, so the first strategic act is measuring your own category on four dimensions.
Concentration. Run your category's recommendation prompts and count how much answer share the top few names hold. A highly concentrated category is a fight for one of a handful of shortlist slots, and displacing an incumbent is the whole strategy. A fragmented category has many partial openings, and breadth of presence pays more than a frontal assault on the leader.
Persona sensitivity. The same recommendation prompt asked as different buyers rewrites the brand list heavily in taste-driven categories like apparel and far less in utility categories like insurance; Prompt and Topic Research covers how to test this. High sensitivity means the aggregate leaderboard hides winnable segments: a challenger can own a persona the leaders never show up for.
The memory-retrieval mix. Run your prompt set with search off, then on, and measure how much the lists change. Across the study, 44.1% of the brands in an average searched answer were absent from the same model's memory-only answer, but the mix varies by category. A category whose answers barely move when search turns on is decided by training-data fame, which is slow to change. A category whose answers rewrite heavily under search is decided by pages and rankings, which you can move in weeks. Model Memory and Live Retrieval is the foundation here.
Commercial value. Weight each prompt class by the revenue it touches. Category shortlist prompts carry the concentration risk; branded prompts carry the accuracy risk; comparison prompts sit between. A category score that ignores which prompts actually precede purchases allocates budget toward visibility that does not convert.
The four starting positions
The category tells you where the game is played. Your position tells you which moves are available.
The unknown brand. The model's memory has no stable association between your name and your category, so memory-mode answers never name you. The first goal is existence: consistent machine-readable facts everywhere models look, per Entities: Becoming a Thing the Machine Knows, and retrieval presence so searched answers can carry you while memory catches up. One caution from the data: Wikipedia and Wikidata presence correlate with AI visibility, but the correlation nearly vanishes once backlink authority is held fixed, so treat encyclopedic presence as a consequence of earned authority, and build the authority.The partial-correlation analysis in What Actually Correlates with AI Search Visibility: backlinks retain 0.077 with the other signals held fixed, Wikipedia and Wikidata drop to roughly zero.
The known-but-misrepresented brand. Models name you, and get your pricing, positioning, or product line wrong. The work is correction, not promotion: one canonical fact set published everywhere machines read, current pages that outdate the stale ones in retrieval, and the governance to keep facts consistent, per Running the Program: Workflow, Team, and Governance and Freshness, Cutoffs, and Temporal Accuracy.
The challenger. An incumbent holds the slot you want. Frontal displacement in a concentrated category is the slowest, most expensive path, so hunt the openings instead: personas the leaders ignore, prompts with no stable incumbent, and sources the models read where the incumbent is absent. Competitive Analysis is the toolkit for finding them.
The leader. Concentration works for you now, and memory is your moat: a competitor cannot un-train the model that learned your name. The exposure is at the edges, personas and new prompt classes where your dominance does not carry, and in retrieval, where a competitor's pages can rewrite searched answers well before they dent your memory position. Defense means monitoring both modes, not celebrating the aggregate number.
Allocation the evidence supports
Across the signals the study tested, backlink authority is the one with a clear independent relationship to AI recommendations, with Reddit discussion volume second and the encyclopedic signals contributing almost nothing on their own. So the authority budget goes to earning links and real community presence, with entity work funded as hygiene rather than as a growth lever. On the retrieval side, 81% of AI top-10 recommendations also rank in Google's organic top 10, so classic SEO spending is still buying AI entry tickets, and cutting it to fund a separate "AI budget" moves money from a working lever to an unproven one.
Where the category scores should change the split: high persona sensitivity shifts budget toward segment-specific pages and prompts; a retrieval-heavy mix shifts it toward content and technical work; a memory-heavy mix shifts it toward PR, links, and community, and extends the timeline you promise anyone.
Sequence by the clocks
Retrieval work shows results in weeks, because it works through pages and indexes that update continuously. Memory work compounds over model training cycles, months at minimum. Sequence accordingly: fund the program with retrieval wins that show up inside a quarter, and run the memory investments underneath them, continuously, on the understanding that they pay later and then keep paying.
Skipping the memory work because it is slow is the common failure. The slow work is the defensible work.
The First 90 Days turns this sequencing into a week-by-week plan with exit criteria.