Does AI reuse the same sources across related questions? 79% of cited domains appear on exactly one of 12 — and the few that span several are the ones that come back

What we measured, and how

Four production runs, pulled by runId (2026-07-28 `jh7bvmn1z0tx0s8cwgk7e8xtqn8bd8fw`, 2026-08-03 `jh70zhz8f7xfe62z8vvxysxze18brfvp`, 2026-08-10 `jh75shn14ccqswchppczr9ychh8c7et1`, 2026-08-17 `jh72qsgx56k5490838rt6fa41h8cnwhh`), 576 answers. Domain = hostname without `www.`, the production sensor's rule; because this is a domain-level cut, Gemini's bare-hostname citations count here (the earlier URL-depth, slug and URLs-per-domain cuts had to exclude it). A domain spans a question if any engine's any sample on that question cited it — set membership, the same rule our self-citation gauge uses. A slot is one (answer, domain) pair: a domain cited by one answer counts once however many of its URLs that answer listed. Span is counted within a run (538 to 571 cited domains per run) and pooled across the four (1,155 distinct domains). Survival: for each week-to-week transition, the share of week-k domains cited again in week k+1, split by week-k span. Earlier cuts of this corpus measured how many domains compete per question and how many pages one domain contributes; this one asks how many questions one domain gets to answer.

The census: one question per domain is the rule

The pooled single-question share is lower than any single run's because four weeks give a domain four chances to land on a second question; even so, 885 of 1,155 domains ever cited were cited on one question only, 149 on two, 58 on three, and 22 on six or more. The head is short: the top 25 domains hold 24.1% of the 5,949 pooled slots, the other 1,130 domains hold the rest.

By engine

No engine carries a generalist bench. ChatGPT and Claude are the most question-specific (82.8% and 81.7% single-question; no Claude-cited domain spans more than 5 of 12). Perplexity is the engine that reuses: its two widest sources, youtube.com and reddit.com, span 12 and 11 questions, and it is the most concentrated at the top (its ten most-cited domains take 20.8% of its slots). Gemini cites the most domains (606) and spreads them the thinnest (top-10 share 12.4%).

Do multi-question domains survive better?

We knew citations that survived one week are far likelier to survive the next, and that domains cited through several URLs persist at 92.7%. Span is a third single-week predictor. For each of the three transitions we took every domain cited in week k and asked whether it was cited at all in week k+1, split by its week-k span; per engine, whether the same engine cited it again.

Per transition, any engine to any engine: 44.4% vs 89.3% (07-28→08-03), 48.5% vs 86.1% (08-03→08-10), 45.1% vs 85.7% (08-10→08-17). The gap holds in all 12 engine-transition cells; the smallest is Claude 08-10→08-17 (61.9% single vs 80.0% multi, n = 84 and 15), the largest Gemini 07-28→08-03 (33.6% vs 85.4%, n = 232 and 48). Correlation, not causation: a domain on several questions is by construction cited more often, and high-citation domains persist for reasons unrelated to breadth. What the cut establishes is that the week-k question count alone separates the flickering tail from the stable core.

The 22 generalists

Only reddit.com is a generalist every week — 11 of 12 questions in all four runs (the one it never reaches: "Is there a one-time AI visibility audit instead of a monthly subscription?"). youtube.com's pooled 12 hides a slide from 9 questions to 4; hubspot.com went 8 → 4 → 3 → 2. Twelve of the 22 are software vendors or agencies in or around the AI-visibility category whose own blogs the engines cite across it (trysight.ai, semrush.com, dageno.ai, amplitude.com, llmpulse.ai, zapier.com, useomnia.com, rankability.com, frase.io, therankmasters.com, tryprofound.com, layer3labs.io); the rest are platforms and publications. Four are user-generated platforms (reddit, youtube, linkedin, medium) — the UGC share we measured before is small in slots but wide in span.

Which questions have the most one-off sources?

The three "best tools" and "alternative to X" questions draw on the shared vendor-listicle pool (36–38% one-off sources); the one-time-audit and local-business questions are answered mostly by sources that appear nowhere else in the panel (63–67%). A page built for one of the specialist-heavy questions competes against one-offs, not against the 22 generalists.

Our own domain

askedabout.com spanned 0, 0, 1 and 3 questions across the four runs: nothing on 07-28 and 08-03; on 08-10 Perplexity cited askedabout.com/compare/best-profound-alternatives on "What is a good Profound alternative for a small business?"; on 08-17 that citation held and ChatGPT added the homepage on "How can I check what ChatGPT says about my business?" (3 of 3 samples) and "Which tools show how ChatGPT, Perplexity and Gemini recommend businesses?" (1 of 3). By this cut's reading we moved from the single-question bucket to the multi-question bucket in the latest run — the bucket that came back 87.0% of the time. The 08-24 run is the test; the self-audit page prints it either way.

Limits

• 12 questions in one category (AI-visibility tooling for small businesses); span is bounded by 12 and the questions are not equidistant — three are near-duplicates about tools, which inflates reuse among vendor listicles.
• Engines were asked through their APIs, not the consumer apps; Gemini's channel returns bare hostnames, which is fine at domain level but means a Gemini "domain" may stand for several pages.
• Set membership: one sample of one engine citing a domain once puts it on that question. Slots are per (answer, domain), so a domain cited by three URLs in one answer counts once.
• Survival rates are per-domain, unweighted; the multi-question buckets are small for ChatGPT and Claude (n = 57 and 40 over three transitions).
• Correlation, not causation, on the survival link.

Why this matters if you care about AI visibility

• A citation is won per question, not per site. 78–80% of cited domains appear on one of 12 related questions in any given week. Being the source for "how do I check what ChatGPT says about my business" earns you nothing, by default, on "what is the cheapest tool" — a page per question is the unit, consistent with the one-URL-per-domain norm.
• Breadth is what persists. Domains on two or more questions came back the next week 87.0% of the time; one-question domains, 46.0%. Whether or not breadth causes it, a second answered question is the visible difference between the core and the flicker.
• There is no generalist gatekeeper to displace. Five or six domains per run span six or more questions, and the top ten hold about 14% of slots. The competition on a specialist-heavy question is other one-offs. Find out which question, if any, you are currently cited on: check which engines name you.

Frequently asked