AI answers concentrate roughly twice as hard as Google organic rankings: in OppAlerts' study across 100 industries, the top five brands held 49% of AI answer share against 24% in Google organic.From the OppAlerts ranking-factors study, 403,000 prompts across 100 industries and 13 models. Method and tables in The Research Behind This Guide. In a channel that concentrated, most of your growth has to come out of a specific competitor's share, so competitive analysis reduces to one question asked precisely: who holds the slot we want, and why do they hold it.
This chapter assumes the machinery from Measuring AI Search Visibility and Building a Visibility Tracking System. Everything below is those tools pointed at other brands.
Benchmark before diagnosing
The benchmark is share of answers, competitor by competitor: across repeated runs of your prompt inventory, what percentage of answers name each brand, at what rank, on which models, for which personas, with search on and off. Repeated runs are not optional. Answers vary across runs, prompts, and time, so a single query says almost nothing about who really holds a slot; the published measurement research reaches the same conclusion, that visibility has to be characterized as a distribution rather than a single observation.Don't Measure Once: Measuring Visibility in AI Search, which shows one-off observations are unreliable and repeated measurement is required to characterize a brand's position.
Two cuts of the benchmark matter more than the aggregate. The per-model cut, because a competitor can own one model's answers and be absent from another's. And the per-persona cut, because in persona-sensitive categories the aggregate leaderboard is an average of several different competitions, and averages hide the one you could win.
Decompose the advantage
A competitor's share comes from model memory, from retrieval, or from both, and the search toggle separates them. Search-off share is training-data fame: years of coverage, links, and discussion compressed into the model's weights. Search-on share is current pages winning current rankings. The decomposition tells you what displacing them costs.
A competitor strong in memory but weak in searched answers is coasting on reputation, and retrieval is already replacing them; the slot is winnable with content and rankings on a timescale of weeks to months. A competitor weak in memory but strong in searched answers is renting their position from the search index, and holding rankings against them is the entire fight. A competitor strong in both gets displaced at the edges first, in personas and prompt classes, never head-on.
You cannot un-train a model that learned a competitor's name. You can outrank the pages retrieval reads tomorrow.
Citation-gap analysis
For searched answers, the models name their sources, and those citations are competitive intelligence. Collect the answers that recommend your competitor and list every cited domain. In the study data, brands' own sites are the largest credited bucket, at least 53.9% of named citations, and the third-party remainder skews toward a short list of publishers per category, with Forbes credited across 51 of 100 industries as the nearest thing to a universal gatekeeper.The citation analysis for the five search-grounded models, summarized in What Actually Correlates with AI Search Visibility, with method in The Research Behind This Guide.
The cited domains where the competitor appears and you do not are a literal work queue: those pages and publishers are demonstrably read and credited in your category's answers. That beats any domain-authority export as a target list, because it is observed behavior rather than a proxy score.
Then check what the models read but do not credit. Reddit is the most retrieved third-party domain in the study while being credited about once per 100 retrievals, so community reputation shapes answers with almost no citation trail. A competitor absent from your category's live threads is exposed in a place their citation reports will never show them; Reviews, Reddit, and Communities covers building presence there legitimately.
Hunting the gaps
Three gap types recur. Unclaimed prompts: flag prompts where answers disagree across runs and models. High variance means no incumbent has consolidated the slot, and unstable slots cost far less to win than settled ones. Ignored personas: in taste-driven categories, leaders often dominate the generic prompt and thin out when the asker changes, and each persona the leaders skip is an entry point. Open sources: publishers and communities the models credit in your category where no competitor has coverage yet.
Rank the gaps by the commercial weight of the prompts behind them, not by how easy they look. A cheap slot on a prompt that never precedes a purchase is decoration.
Defending versus attacking
The same instruments run in both directions. Attacking, you sample the leader's territory for variance and thin personas. Defending, you watch your own: rising run-to-run variance on prompts you used to own is the earliest signal that your slot is loosening, visible before any share actually moves. Defenders should also reread the category's credited-source list quarterly, because a challenger building coverage across those publishers is announcing the attack in public, inside the models' own reading list.
Either way, rerun the full benchmark on a cadence, monthly is enough for most categories, and log every run, so that when share moves you can connect the movement to what changed: their pages, your pages, or the model itself. Running the Program: Workflow, Team, and Governance makes that logging a standing procedure.