AI search engines choose sources by mixing two jobs: what the model already “knows” from training, and what it can fetch and quote for this answer. ChatGPT, Claude, Gemini, Perplexity, and Google AI features do not share one picker. On 13 Oct 2026 this URL still showed an April layout, 4 Search Console impressions / 0 clicks (28d), and was indexed (last crawl 9 Oct 2026). Keep Search Console next to the checker.
First published 12 Apr 2026. Full rewrite. April copy is labeled history only — do not reprint its unauditable “thousands of recommendations” tier table.
What we tested (13 Oct 2026)
Unique data is this URL’s Search Console row, URL Inspection, live HTML, and today’s Google SERP — not a roundup’s paraphrase of model weights, and not a student guide on how to cite ChatGPT in a paper.
- This URL, live before rewrite: HTTP 200, canonical
https://scope.online/learn/how-ai-search-engines-choose-sources. Layout date still 12 Apr 2026. Body ~180,573 bytes of April shell. No datedWhat we tested (DATE)box. April framing claimed a cross-platform “authority hierarchy” from “thousands” of business recommendations without a primary sample, date range, or engine log you can audit on this page. - URL Inspection, 13 Oct 2026: coverage Submitted and indexed. Last crawl 9 Oct 2026 (
2026-10-09T21:30:21Z).robots.txtALLOWED. Fetch SUCCESSFUL. User canonical = Google canonical = this URL. Rich results: Breadcrumbs only. Indexed ≠ ranking for the H1; that crawl is still the April shell until this rewrite is recrawled. - Search Console, this URL, LAST_28_DAYS (13 Oct 2026 pull): 4 impressions, 0 clicks, CTR 0, average position 77. Page EQUALS query report returned one named row:
what platforms analyze which sources ai models prefer for answers?— 4 impressions / 0 clicks / pos 77. That is B2B operator intent (who measures source preference), not an APA homework query. Page-level average position is not a durable H1 rank on four impressions. - Search Console, site totals, 13 Oct 2026 pull: 31 clicks, 4,988 impressions, CTR ~0.62%, average position ~25.3. Site totals are truth; page breakdowns under-count.
- Google SERP, desktop US, 13 Oct 2026 (DataForSEO live/regular):
how ai search engines choose sources— Scope absent. Page one: operator explainers (OxAI Studio, LinkedIn pulse), Profound optimize-for-AI-search article, Quora operator threads, Visiblie RAG explainer, Melbourne SEO Services, Single Grain citation SEO. B2B marketer / AEO operator intent — owner-SERP pass.how do AI search engines select sources— Scope absent. Cheers.tech GEO academy, agency “rank in AI search” guides, AIO source posts, Coveo definition. Same operator job.how chatgpt chooses sources— Scope absent. Zapier, independent network-trace posts, Morphic, Conductor academy, Reddit. Still operator/B2B — not student citation style.ai citation sourcesalone — wrong job (library APA / “cite AI in essays”). Do not retitle this page into that SERP.- First multi-query DFS batch hit 40101 Internal SE Server Error on two H1 variants; those were retried one-at-a-time and returned organic. Record transient SE errors — do not treat them as “0 results = we rank #1.”
- Answer engines this run: chatgpt.com / perplexity.ai are not treated as citation sources when blocked or shell-only. Record blocked / shell, never invent citations.
- Sitemap: live
/sitemap.xmllists this<loc>among 341 URLs. Sitemap inclusion did not create H1 ownership.
The gap is not “missing a sixth fake study of every engine’s weights.” It is an April shell that is indexed but almost invisible on Google for the H1, while page one already answers the job with vendor guides and operator essays.
How do AI search engines choose sources in plain language?
Strip the vendor slogans. For a buyer prompt, an engine roughly does some mix of:
| Stage | Question the system is answering | What fails look like for you |
|---|---|---|
| Gate | Can anything fetch a clear public page? | Soft 403, JS-empty first HTML, noindex, wrong canonical |
| Retrieve | Which URLs or passages enter the candidate set? | You never appear in the fan-out results the model sees |
| Extract | Can a model quote a fact without the rest of the site? | Vague homepage prose; proof buried below the fold |
| Entity | Is “you” the same business across pages and listings? | Homonyms, NAP drift, two conflicting About stories |
| Compose | Mention, cite a URL, recommend, or omit? | Competitor named; you omitted; or cite lands on /about |
Deeper stage framing lives on AI search ranking factors. Citation tactics live on how to get cited in AI search. This page’s job is the source-selection split across engines — why ChatGPT and Perplexity can disagree on the same day.
AI source selection is not:
- One leaked universal ranking formula you can buy as a badge.
- A reason to pause Google SEO (crawls, links, indexable pages).
- A special AI-only side file that tickets you into ChatGPT. Google Search Central (AI features, confirmed 2 Sep 2026) still centers pages that are indexed and snippet-eligible.
- Student instructions for citing ChatGPT in a term paper (wrong job for this H1).
Training memory vs live retrieval — why the split matters
Two modes shape almost every “why did they cite them?” post-mortem:
Training-heavy answers (model without a browse/search tool, or when the product chooses not to search): the system leans on patterns and entities that showed up often in pretraining and post-training data. Fresh blog posts from last week may not help. Editorial mentions, category-defining docs, and long-lived entity pages matter more — and move slowly.
Retrieval-heavy answers (Perplexity-style always-on search, ChatGPT with search, Claude with web tools, Gemini grounded on Google’s index, AI Overviews / AI Mode): the system runs a fan-out of queries, pulls candidate pages, then composes. Current public pages, extractable answers, and consensus across multiple sources matter more this week.
Most products blur the line. A single ChatGPT session can answer from memory, then search, then mix both. That is why a frozen prompt panel beats a one-shot anecdote.
| Mode | What tends to matter | What does not save you |
|---|---|---|
| Training-heavy | Repeated entity mentions in durable sources; clear category neighbors | A same-day title tweak with no off-site repeat |
| Retrieval-heavy | Fetchable money URL, first-HTML answer, multi-source agreement | A long essay that never states the fact in plain language |
| Google AI features | Index + snippet eligibility on Google | Assuming a ChatGPT-only checklist covers AI Overviews |
Practical rule: keep Google SEO running; add an assistant panel. Agencies folding this into retainers keep the SEO line — see how agencies add AI visibility to SEO.
How does source selection differ by engine family?
Do not collapse ChatGPT ≠ Claude ≠ Gemini ≠ Perplexity ≠ AI Overviews into one tactic list. Patterns below are operator heuristics from public product behavior and today’s SERP book — not invented win-rate percentages.
ChatGPT (OpenAI)
- Without search: leans on training memory and post-training preferences. Category leaders and frequently compared brands show up more often.
- With search / browse: runs web retrieval; lists, docs, UGC, and YouTube can enter the mix more than on some other engines. Cited URLs may not match Google’s #1 for the same string.
- Operator miss: treating a memory-mode omission as a “broken sitemap” problem, or treating one browse citation as proof you “rank on ChatGPT.”
Perplexity
- Search-first composition: nearly every answer surfaces visible citations from live retrieval.
- Operator pattern: pages that already answer the question cleanly and rank in conventional search often appear; niche authoritative pages can sit next to larger domains when the extract is better.
- Operator miss: writing only brand-voice essays with no quotable fact, then expecting a citation.
Claude (Anthropic)
- Base model: training-dependent when tools are off.
- With web tools: tends to be citation-conservative — prefers higher-confidence sources over a pile of weak ones. May refuse a specific local recommendation when it cannot verify.
- Operator miss: copying a ChatGPT UGC-heavy playbook and assuming Claude will mirror it. Claude’s public citation mix has historically under-weighted some community sources relative to ChatGPT.
Gemini and Google AI features
- Grounded on Google’s world: index quality, entity clarity, and Business Profile / Knowledge signals matter more than on a pure chat product.
- AI Overviews / AI Mode: still require pages Google can index and snippet. They are not the same product as ChatGPT chat, even when the brand question looks similar.
- Operator miss: shipping a ChatGPT-only checklist and declaring “Google is handled.”
Cross-check table (use as a log header, not a score)
| Engine family | Default bias to watch | First fix when you lose |
|---|---|---|
| ChatGPT | Memory vs search mode confusion | Freeze prompts; note whether search ran; check cited URL |
| Perplexity | Live citation set | Extractable answer + conventional discoverability |
| Claude | Conservative verification | Entity clarity + authoritative corroboration |
| Gemini / AIO / AI Mode | Google index + entity | Search Console coverage + first HTML clarity |
Related engine playbooks: how Claude ranks businesses (keep recommends language where that URL’s H1 requires it), which AI platforms cite small businesses most.
Who owns “how AI search engines choose sources” on Google today?
Condensed from the 13 Oct 2026 US desktop pulls:
| Pattern on page one | Example domains | Copy this | Do not copy |
|---|---|---|---|
| Operator mechanism essay | OxAI Studio, Visiblie, agency GEO posts | Training vs retrieval + stage language | Invented multi-thousand-query “studies” without a sample |
| Vendor guide | Profound | Clear optimize-for-AI-search steps next to SEO | Cloning a competitor’s full product suite as your page |
| Citation-SEO pitch | Single Grain and similar | Off-site consensus as a real lever | Claiming one link-building product owns every engine |
| ChatGPT-specific | Zapier, Morphic, Conductor academy | Product-mode split (memory vs browse) | Student APA “how to cite ChatGPT” titles |
| Forums | Quora, Reddit | Real buyer wording | Treating a thread as proof of your rank |
Operator read: an indexed April essay with 4 impressions at average position 77 is not “we own source selection.” Before arguing another tier pyramid, clear the first answer, date a test, and log engines separately. After that, off-site consensus (lists, docs, independent pages) moves assistants more than a fourth synonym for the same H1.
What we will re-check after deploy: HTTP 200 with needle What we tested (13 Oct 2026), layout date 13 Oct 2026, sitemap still containing this loc, and a later Inspection crawl of the new HTML. Indexation lag is normal. Do not extra-submit the sitemap the same morning as a Learn rewrite unless last download is already stale on a separate technical gate.
Worked example: B2B SaaS “best X for Y” prompt
Setup. A 12-person B2B SaaS sells a workflow tool. Marketing freezes five prompts, including “best [category] for mid-market teams 2026.”
Week 0 log (honest):
| Engine | Mode observed | Outcome | Cited URL if any |
|---|---|---|---|
| ChatGPT | Search on | Competitor recommended | Competitor compare page |
| Perplexity | Always-on citations | Competitor + two listicles | Listicle URLs; you omitted |
| Claude | Web tools on | Soft deferral / category leaders only | Docs-style domains |
| Gemini | Grounded | Competitor + Google-favored review | Review hub |
What they do (30 days, not theater):
- Money URL first HTML: 60-word answer naming who they serve, one proof metric with date, and a clear price or plan boundary.
- Search Console: confirm the money URL is indexed; fix accidental noindex on the compare page.
- Same proof sentence repeated on About, compare, and one independent listing they control (create/claim — not cold bulk email).
- Re-run the same five prompts; log omit/mention/cite/recommend again.
What “better” looks like: not a global share-of-voice fantasy. One engine moves omit → mention, or cite shifts from a random blog to the money URL. Google clicks on branded and category queries stay on their own panel.
Counter-example: the unauditable tier pyramid
April-style pages often ship a Tier 1–4 source hierarchy (“almost always cited…”) as if it were a multi-engine census. That pattern fails three ways:
- No sample you can audit — no date range, no prompt set, no engine list, no raw log.
- Collapses engines — Yelp-heavy local heuristics get pasted onto B2B SaaS prompts where G2/docs/listicles dominate instead.
- Invites the wrong SERP — titles that sound like “AI citation sources” drift into student APA results.
If you need a directory checklist for local operators, keep it on a local-specific URL and label it as NAP/listing hygiene — not as a universal model of how every engine chooses sources. Prefer stage language (gate → retrieve → extract → entity → compose) over fake precision tiers.
30-minute operator checklist
Use this when someone asks “why does ChatGPT cite them and not us?”
- Freeze five prompts buyers actually ask. Write them down. Do not change wording mid-test.
- Run each engine once. Note whether search/browse ran. Screenshot or paste outcomes into a sheet: omit / mention / cite / recommend + cited URL.
- Curl your money URL. Confirm HTTP 200, self-canonical, and a 40–80 word answer in first HTML (view-source, not only the painted page).
- Open Search Console for that URL. Coverage state and last crawl beat a vibes debate.
- Diff the cited competitor URL. What fact did they put in the first screen that you buried?
- One off-site repeat. Same entity sentence on a listing or doc you legitimately control. No purchased-list email.
- Re-run only the failed prompts after the fix ships live (GET 200 + visible change). Log the delta. Stop after one cycle the same day — recrawl lag is real.
Optional product path: AI visibility checker, then Scope MCP for agents if your team works in Claude/Cursor. Compare vendors with honest non-claims on Scope vs Profound.
How does this sit next to Google SEO?
| Signal | Google SEO | Assistant source selection |
|---|---|---|
| Primary log | Search Console | Prompt panel per engine |
| Hard gate | Index + snippet eligibility | Public fetch when browsing is used |
| “Win” | Click, impression, position | Named or cited in the answer |
| Common miss | Thin title, noindex, cannibal | Homonym, empty first HTML, no consensus |
| Off-site mix | Links, brand queries | Lists, docs, UGC — engine-dependent |
You can rank on Google for a related query and still be omitted when someone asks ChatGPT who to use. You can be mentioned without a citation, or cited on a URL that is not your money page. Measure both panels. Do not collapse them into one blended “AI SEO %.”
Pair frame for stakeholders: GEO vs SEO. Definition of the measurement language: what is AI visibility.
FAQ
Q: Do all AI search engines use the same sources?
A: No. Training-heavy answers lean on long-lived entity and category patterns. Retrieval-heavy answers pull live candidates and compose. ChatGPT, Claude, Gemini, Perplexity, and Google AI features weight those paths differently. Log them separately.
Q: Is this the same as “how to cite ChatGPT” for school?
A: No. That is a student attribution job (APA/MLA library guides). This page is about how answer engines select web sources when recommending or explaining businesses. Do not merge the H1s.
Q: Does better Google ranking automatically mean more AI citations?
A: Overlap exists — especially for Gemini and AI features that need indexable pages, and for retrieval modes that fan out like search. It is not automatic. You can win Google and still be omitted on ChatGPT the same week.
Q: Should I optimize separately for each engine?
A: Share foundations (fetchable money URL, extractable answer, consistent entity, honest facts, healthy Search Console). Specialize the log and the off-site mix. Do not maintain four unrelated content farms.
Q: What is the fastest fix when one engine cites a competitor?
A: Read the cited URL. Match the missing extractable fact on your money page first HTML, confirm indexation, re-test the same prompt, then earn one off-site corroboration. Skip a full site redesign on day one.
Q: Do I need llms.txt or special AI schema to be chosen as a source?
A: Treat exotic side files as optional plumbing. Google’s public AI-features guidance still centers indexed, snippet-eligible pages. Allowing known retrieval bots and clear first HTML matter more than a new file fad.
Q: Can I publish a tier list of “sources AI always cites”?
A: Only with a dated sample, engine list, and prompt set on the same page. Unauditable tier pyramids are the failure mode this rewrite removes. Prefer stage checklists and per-engine logs.
Q: Where should a marketing team start today?
A: Run the checker, freeze five prompts, fix the money URL gate and first 80 words, keep Google SEO running, and read content strategy for AI search for the long-form bar.
Sources and method notes
- Search Console site + page + query pulls for
sc-domain:scope.online, LAST_28_DAYS, 13 Oct 2026 (Markifact). - URL Inspection for this URL, 13 Oct 2026 (Markifact).
- DataForSEO Google organic live/regular, desktop US, 13 Oct 2026 for
how ai search engines choose sources, sister select/cite queries,how chatgpt chooses sources, and the wrong-job controlai citation sources. - Google Search Central documentation on AI features in Search (indexed + snippet-eligible pages) — https://developers.google.com/search/docs/appearance/ai-features
- Live HTML curl of this URL before rewrite (April layout, ~180 KB, no dated test box).
April 2026 first publish remains the historical start date in the body above. Header date / updatedAt are the ship day of this rewrite so the Learn layout shows 13 Oct 2026.