ChatGPT-5 has been out for six weeks. We’ve been running citation tracking on 12 client portfolios across supplement, SaaS, DTC luggage, and home services categories — measuring whether each brand gets mentioned by name when a category-defining query is run on -4 vs -5.

Here’s what the data shows. It’s directional — the model is still being tuned and our sample is small — but the patterns are clean enough to act on.

Headline finding: -5 is more conservative with brand mentions

For the same query, ChatGPT-5 mentions a specific brand by name 23% less often than -4, on average across our 12 portfolios.

When -5 does name a brand, it tends to be the most-mentioned one in its training corpus by a wider margin. The mid-tier mentions that -4 used to surface (the 2nd-and-3rd-ranked brands in the category) are getting compressed.

In practical terms: if you were the 1st or 2nd most-cited brand in your category on -4, you’re probably still mentioned (maybe more often). If you were the 3rd–5th most cited, you’re probably down 30–50% in citation rate.

Wikipedia signals are more heavily weighted

The biggest single factor predicting whether a brand got cited on -5 was whether the brand had a Wikipedia article. In our sample:

  • Brands with Wikipedia articles: cited in 71% of relevant test queries on -5 (vs 64% on -4).
  • Brands without: cited in 38% of relevant test queries on -5 (vs 51% on -4).

The drop for non-Wikipedia brands is the killer number. Wikipedia presence used to be a "nice to have" for AEO. With -5, it’s looking like a soft requirement for consistent citation.

This tracks with what Anthropic and Google have publicly said about RAG: knowledge-graph-anchored entities are getting weighted more heavily than free-floating training mentions. Wikipedia is the largest open knowledge graph; getting an entry is increasingly the price of admission.

Reddit signals: still strong, but more discriminating

Reddit mentions are still highly predictive of citation rate on -5 — but the quality of the mention matters more.

A high-engagement Reddit thread with 500+ upvotes and substantive discussion is roughly 3x more likely to feed a citation than a thread with 5–20 upvotes mentioning the brand once. On -4, the ratio was closer to 1.6x.

This suggests -5 is using engagement metrics to weight Reddit content during retrieval, not just training. Brands that have aged Reddit accounts contributing to high-engagement threads will keep winning. Brands that paid for low-effort Reddit comment farms will see those signals decay.

PR placements: more durable than expected

We worried PR-driven citations would be the most volatile signal across the version change. They’ve been the most stable.

A brand that got placed in 3+ tier-1 publications (Forbes, Inc., Healthline, Wired) in 2024–2025 maintained near-identical citation rates between -4 and -5. The signal is durable across model versions because the publications themselves are weighted heavily in retrieval, regardless of which model version is doing the retrieval.

Read this as: PR investment compounds across model upgrades better than community signals do. If you’re allocating budget, it’s still the highest-leverage long-term lever.

What’s gone

Three things stopped working as well on -5:

  1. Anchor-text-only mentions. A brand name embedded in a URL with no surrounding context used to occasionally surface in -4. -5 mostly ignores these.
  2. Press release wire citations. Wire-distributed press releases with the brand mention contributed to -4 citation rates. -5 appears to discount these heavily.
  3. Self-published "best of" lists. Brand-owned blog posts ranking themselves as "top X" never surfaced consistently, but they had a non-zero contribution on -4. On -5, near zero.

What we’re shipping in response

Three program changes for clients running AEO with us:

  1. Wikipedia eligibility audit for every client. If a brand is plausibly notable enough for an article and doesn’t have one, we’re adding the work to scope. The effort is significant (real notability requires real independent press) but the AEO ROI is now unambiguous.
  2. Reddit engagement quality threshold. We’re scoring Reddit threads by engagement before placing — only contributing to threads with 200+ upvote velocity, because lower-engagement threads aren’t paying out at -5 like they did at -4.
  3. Doubling tier-1 PR cadence for clients near tier-1 thresholds. The cross-model durability of PR makes it the cheapest insurance against future model version changes. Clients who can shoulder the cost are getting pushed to ship 1+ tier-1 placement per month.
The TL;DR for any AEO program: Wikipedia entity work + tier-1 PR + high-engagement Reddit. Skip the rest.

We’ll re-run the citation tracking at three months and six months and publish updates. If you want us to run the same tracking for your portfolio, book an audit and ask for the AEO citation report — we’ll show you where you sit on -5 today.