Interesting thing I'm watching: GPT-5 seems to handle brand probes differently than GPT-4o did
Petr here. Not a product pitch — just something I've been watching across the domains we monitor.
Since the GPT-5 rollout, brand-probe queries ('what does [brand] do', 'is [brand] worth it') are returning noticeably longer responses with more source-citing behavior, even without web search turned on. The training corpus is clearly richer.
Three patterns I'm seeing across roughly 200 domains:
- Brands that had broad entity presence (10+ external pages mentioning them) before the GPT-5 cutoff are holding or improving. Makes sense — more data averaged.
- Brands that relied on a single HN front-page spike for their citations are losing position. Spike-only patterns don't compound across retraining cycles.
- Hallucination rates on pricing/features are slightly down vs GPT-4o. Not dramatically, but measurable.
None of this changes the playbook — broad entity presence, comparison pages, review velocity. But the 'one good piece of content' strategy seems riskier now that retraining cycles are picking up speed.
Anyone else noticing this? Curious what patterns you're seeing in your own data.
4 replies
- Nora
The pricing/features hallucination rate being down in GPT-5 is interesting. Are you seeing the same in Claude? And is the improvement mostly for well-documented products or across the board?
- Marcus B.
Curious about the sample shape here. 200 domains is a meaningful n, but are the three patterns weighted by domain or by query? If you ran ~50 queries per domain and 60% are buyer-intent vs 40% brand probes, the 'broad entity wins' pattern might be category-specific. We saw something similar after the GPT-4o → 4o-mini routing shift — the apparent 'broad entity boost' was really a long-tail compression artifact.
- Dave A.
The 'hallucinations on pricing slightly down' point is the one I'd want broken out. Down from what baseline, and is it across all engines or just ChatGPT? My hunch is any improvement is from search-on becoming default, not from anything intrinsic to GPT-5. If you have the pre/post search-on numbers separately, that'd settle it.
The spike-only pattern losing position is something I've been worried about. Got an HN front-page about 6 months ago that spiked my Perplexity citations hard. If the compounding depends on having pre-built breadth, my breadth at that time was probably thin. Makes sense that the spike didn't fully sustain.