What's New in the AI Score: Prices, Hedging, Freshness
We updated the scoring model based on the largest study of AI citations to date — 252,000 runs across 6 models. Here's what changed, why, and what honestly stayed the same.
TL;DR
- Explicit prices and facts in your copy now add to the score; hedged wording ('may', 'possibly') subtracts
- Freshness carries more weight: fresh dates are a citation gatekeeper, not a bonus
- Crawlability distinguishes three bot classes: training, search, agent
- ChatGPT monitoring moved to the new model generation (GPT-5.6)
- Unchanged: llms.txt is still zero, FAQ schema still works (41% vs 24%)
Why we updated the score
SIGIR 2026 published "What Gets Cited" (arXiv:2605.25517) — 252,000 controlled runs across six LLMs. It's the largest dataset to date on what AI models actually pick to cite. The findings are specific: gatekeeper factors (orders-of-magnitude effects) are topical relevance, explicit prices, and fresh dates. Secondary factors (2–600x) are spec completeness and unhedged claims. And "clean markup by itself is insignificant."
The context shifted in parallel: OpenAI released GPT-5.6, and Reddit's share of ChatGPT citations dropped from ~60% to ~10% over the year after a retrieval overhaul. A scoring model built on last year's data was starting to drift from reality — so we fixed it.
This is the honest build-in-public part: a scoring model is a hypothesis you keep checking against data. When the data says 'you're overweighting X and underweighting Y,' the right move is to reweight — not defend the old weights.
What changed
1. Explicit prices and specs — a bonus
If a page states a concrete price or specification as plain text, the content part of the score now accounts for it. Per the SIGIR data, explicit prices are one of the citation gatekeepers: pages that have them get cited orders of magnitude more often, all else equal.
2. Hedged wording — a penalty
Copy overloaded with 'may', 'could', 'possibly' (and their Russian equivalents), with not a single concrete fact or price, now gets flagged and loses content points. In the study, the effect of hedging on citations was measured in multiples — from 2x to 600x depending on the factor.
3. Freshness carries more weight
Visible dates and dateModified were already in the score, but the study showed fresh dates are a gatekeeper, not a nice-to-have. We raised the freshness weight in the deep audit. Extra context: OpenAI started receiving real-time Cloudflare signals, so freshness in ChatGPT now works on the scale of hours, not days.
4. Three AI crawler classes in crawlability
The access check now distinguishes training bots (GPTBot, ClaudeBot), search bots (OAI-SearchBot, Claude-SearchBot), and agent bots (ChatGPT-User, Claude-User). Agent bots are reported as informational for now, with no effect on points: these are on-the-fly fetches, not an index. Cloudflare will start blocking bots by these same classes in September — distinguishing them is becoming mandatory.
5. ChatGPT monitoring — on the new model generation
OpenAI released GPT-5.6 on July 9. Our ChatGPT monitor moved to the new generation — citation behavior shifts noticeably between model generations (we saw this with Grok), so monitoring on an outdated model means showing you yesterday's picture.
The practical effect: your score may move by a few points. If your copy is long but hedged, with not a single price, you'll see a new issue in your report with a concrete recommendation on what to rewrite.
What did NOT change
llms.txt — still near zero
A 300k-domain study: zero correlation with citations. Google officially doesn't use it. We've kept its weight minimal for a long time and we're keeping it that way: hygiene for AI agents, not a lever. If someone sells you llms.txt as an 'AI visibility secret' — that's not true.
FAQ schema — still works
41% vs 24% citation rate in a controlled study on Perplexity. One of the few changes with a measured effect. Weight unchanged.
Schema.org — necessary, but not decisive
Markup identifies your business to AI — without it the model doesn't know who you are. But 'clean markup by itself is insignificant' is a direct finding of the study. Its weight in our deep audit was already moderate; we've aligned the wording across the site: markup is the baseline, content earns the citations.
Reddit — we never recommended it anyway
After OpenAI's retrieval adjustment, Reddit's share of ChatGPT citations dropped from ~60% to ~10%. 'Post on Reddit for ChatGPT' was never in our recommendations — we double-checked every fix playbook. For Google AI Overviews, community mentions still work.
Honest limits
- Price and hedging detection are heuristics for English and Russian. On other languages the signal simply won't fire (neutral, never a penalty).
- The score measures machine-readability and alignment with citation factors — it does not guarantee citations. The real picture comes from monitoring: live queries against ChatGPT, Gemini, Grok, and Claude.
- The study measured specific models at a specific point in time. Models change — that's why we monitor on the current generation and will reweight again when the data says so.
Sources
Free AI Readiness Check
See your GEO score before reading
Instant score · No email required · Checks ChatGPT, Gemini, Grok signals