GEO Evidence Changed in 2026: What Holds and What Doesn’t

Short answer: adding statistics and expert quotes so AI cites you more, the most repeated GEO shortcut, did not hold up on current AI engines in a 2026 replication. What gained support is plainer: pages AI can find, accurate facts on your own website, and measurement that repeats the same questions. For a small business, that means less money on tricks and more attention on the basics.

When the evidence changes, whoever gave the old advice should say so, and that includes us. Here is what changed between June and September 2026, with the source next to each claim, and what we corrected in our own guides.

How to read the sources: Official documentation (Google, Apple, OpenAI, Anthropic) describes how a product works. Preprints are research papers posted before peer review. Vendor data comes from companies that sell SEO or AI tracking tools. We label each one. When a recommendation goes beyond the sources, it is our editorial judgment, and we say so.

The short version: what holds and what lost support

What did the Princeton GEO study claim, and why doesn't it hold now?

Most "add statistics and quotes" advice traces back to the 2023 Princeton GEO benchmark (Aggarwal et al.). Its headline, "up to 40%" more visibility, became the industry's standard promise, and we repeated it too.

Two 2026 papers changed how to read it:

• The replication. Bajemon and Rochet (arXiv 2609.07559, September 2026, preprint) re-tested adding quotations, statistics and source citations on 10 current engine families, including OpenAI gpt-5.x models re-run in July 2026, using paired edits with the amount of added text held constant. None of the three edits moved citations; the effect was indistinguishable from zero.
• The 45-study review. Martinez (arXiv 2607.14035, July 2026, preprint) explains that the original gains measured how much of an answer was credited to a source the model already had in front of it, not whether a page gets found or brings traffic. No reviewed technique had a stable, long-term effect on being discovered across platforms; relevance to the question and position in the model's context were the levers that reproduced. The review also warns that rewriting pages to chase citations can make them harder to retrieve.

What these papers don't show

Both are preprints, and this is one replication. Neither says statistics and quotes are useless: the review itself recommends "relevant, verifiable, dated, and properly attributed evidence." We keep using sourced, dated figures because they help readers check what we say, not as a promise of citations. What lost support is the "GEO rewrite" that only adds statistics and quotes.

What gained support in 2026?

First an engine has to find your page; then accuracy and clarity shape how it describes your business.

• Being found for the exact question. In the 45-study review, relevance to the question was the most consistently reproduced lever. Well-known brands tend to win broad questions (Ranqo, a vendor-authored preprint), so make sure your important pages are indexed and answer the specific questions your customers ask, not only the broad category.
• Your own website as the source of truth. Brand-owned sites were 57% of AI citations in Profound's analysis (vendor data; the share varies by platform and industry). Trackers also reported that in August ChatGPT's app began searching official sites directly (vendor data; OpenAI hasn't confirmed it).
• Accurate prices and dates. In Profound's study of AI errors (vendor data), pricing was 12% of the claims checked but 24% of the mistakes, and 54% of brands had at least one false claim that cited their own website. An experiment cited in the 45-study review (Vishwakarma et al., preprint) found that explicit prices and recent dates helped as secondary factors when AI chose sources. Show an "Updated" date only after a real review; we found no published evidence of a "refresh multiplier."
• Your home page's first screen. Since May, about 62–63% of ChatGPT referrals land on the home page, up from 26–29% (Similarweb, vendor data). That first screen is where many AI visitors meet your business.
• Structure on pages AI already finds. CITECHOICE (arXiv 2609.15164, preprint) found that tables and lists added about half a citation per answer on pages that had already been retrieved, but the effect on getting cited at all was inconclusive.
• Pages an AI agent can use. In a controlled test with small prototype stores and older models, a page design built for AI agents raised their task success from 49.3% to 89.3% (arXiv 2607.12056, preprint). Anthropic's browser-use tool works through a page's accessibility tree and screenshots, so clearly labeled forms and buttons matter.

What lost support?

Google says it doesn't use some of these tactics; for the rest, new data found no effect or a new risk.

• llms.txt, chunking and special schema as Google tactics. Google's AI optimization guide, updated July 10, 2026, says Google Search ignores llms.txt, doesn't require splitting content into chunks and needs no special schema. It points instead to original, non-commodity content, quality images and video, Merchant Center and Google Business Profile.
• Schema as a citation lever. In Ahrefs' May 2026 test (vendor data), adding schema produced no significant citation lift. Its value is eligibility for Google features, and Google requires markup to match visible, truthful content (structured data guidelines).
• Seeding Reddit for ChatGPT. Reddit's share of ChatGPT citations fell from about 3.8% to 0.5% in mid-August 2026 (Promptwatch; Otterly saw a similar drop). Both are vendor data, and Promptwatch says to trust the direction and timing, not the decimals. OpenAI hasn't confirmed a cause, and a similar drop in September 2025 reversed within weeks. Reddit is still cited in AI Overviews, AI Mode and Perplexity.
• Self-ranked "best X" lists. In Lily Ray's test (June 2026, one practitioner study), AI Overviews cited brands' own "best of" pages but left the brand out of the recommendation 224 of 323 times.
• Fabricated mentions. Google's spam policies (updated August 28, 2026) now cover attempts to manipulate generative AI responses in Google Search, and its AI guide lists chasing inauthentic mentions among the things to skip.
• Mass templated rewrites. Beyond the review's warning about retrieval, researchers are building detectors for GEO-rewritten pages (preprints; no engine is known to use them yet).
• Content Signals and aipref lines in robots.txt. No known effect on Google, OpenAI, Perplexity or Anthropic. Google's John Mueller said Content Signals have "no effects whatsoever for any crawler or LLM" (Search Engine Roundtable).

One more myth: there was no "August 2026 core update." Google's status dashboard shows no core update since May, only spam updates in June, in August and one that began September 24.

What can you measure now, and what can't you?

Google and Microsoft now give site owners free, official data on AI search:

• Search Console's generative AI performance report (all sites since August 31, 2026; Google help). It counts how often links to your site appeared in AI Overviews and AI Mode, by page, country, device and date. It shows impressions only: no clicks, no queries and no split between the two features. Those impressions are already included in the Web search type of the Performance report, so don't add them twice. Low-volume sites may not see the report yet, which isn't the same as zero visibility.
• The "Search generative AI" setting (Google help). Include is the default; Exclude removes your site from AI Overviews, AI Mode and Discover's AI features within one to two days. It doesn't stop AI training, which is controlled separately with Google-Extended. In our view, most small businesses should keep Include.
• GA4's "AI Assistant" channel (since May 13, 2026; GA4 help). Traffic from sources like ChatGPT, Gemini, Deepseek, Copilot and Grok gets its own default channel. Clicks from AI Overviews and AI Mode are not in it; they stay in Organic Search. AI visits that arrive without a referrer show up as Direct, so treat the channel as a minimum, not a total.
• Bing Webmaster Tools AI Performance (Bing). Citation data for Copilot and Bing's AI answers, with new views in preview since June.

ChatGPT, Claude, the Gemini app and Siri don't give site owners a citation or visibility report, so you sample: ask your customers' real questions several times on each platform and compare averages. In studies covered by the 45-study review, repeating the identical query changed 9–28% of source decisions, so a single check proves little. To decide where to spend effort, usage helps: in August 2026, Similarweb's website-only data put ChatGPT at 55.5% of generative AI traffic, Gemini at 25.6%, Claude at 9.3% and Perplexity at 0.9% (The Decoder; Google's AI Overviews and AI Mode aren't included).

Then keep observations apart from business results. A mention or a citation is a visibility observation; a visit, a call, a form and a new customer are separate steps, and many can't be traced back to a specific AI answer. How to appear in ChatGPT shows a simple observation log, and our GEO guide explains how to set a baseline and measure progress without over-attributing.

Why does Siri in Spanish matter if your customers speak Spanish?

Siri now answers questions with information from across the web, and Apple says Spanish is coming in October. Apple launched this beta on September 14, 2026, in iOS, iPadOS and macOS 27 (Apple); for now it works only in English. Once Spanish arrives, your customers will be able to ask Siri about businesses like yours in their own language.

Three details matter:

• Siri reads the web through Applebot, not Googlebot. Apple says its models were built in collaboration with Google, but showing up in Google's AI answers doesn't mean showing up in Siri's. Make sure your robots.txt allows Applebot.
• nosnippet opts you out. Apple's Applebot documentation treats the nosnippet meta tag as an opt-out from Siri's web answers. Applebot-Extended only controls AI training.
• There's no Siri visibility report. Once Spanish launches, ask Siri your customers' questions in Spanish and write down the answers.

What should a small business do now?

This checklist is our editorial judgment, based on the sources above. It isn't a ranking formula and doesn't guarantee a mention.

1. 1

   Check your Google AI setting

   In Search Console, open Settings → Search generative AI and confirm it says Include (if it says Inherit, check that the parent property is set to Include).

2. 2

   Save a baseline

   In the generative AI report, note the impressions for your most important pages. If the report doesn't appear, write "low volume," not "zero."

3. 3

   Check GA4's AI Assistant channel

   Confirm it appears in Traffic acquisition, then look in Referral for AI sources it misses.

4. 4

   Make your prices and facts match

   Put prices or ranges, hours, services and service area on your own pages, and keep them identical on Google Business Profile, directories and social profiles. Example: a fictional café whose website says it opens at 7 a.m. while its profile says 8 a.m. gives AI two answers; confirm the correct hours and fix both.

5. 5

   Review your home page's first screen

   Say what you do, for whom, where, what it costs if you can publish it, and how to contact or book you.

6. 6

   Skip the shortcuts

   No rewrites that only add statistics, no self-ranked lists, no seeded Reddit posts, no bought mentions.

7. 7

   Measure with the same questions

   Keep a fixed list of real customer questions in English and Spanish, ask each one at least three times per platform at regular intervals, and compare averages. Note model launches, and keep mentions separate from visits, inquiries and customers.

8. 8

   Prepare for Siri in Spanish

   Allow Applebot and remove nosnippet from pages you want Siri to use.

What we corrected on our own site

Alongside this article, we corrected our published guides. The main changes:

• Princeton figures. How AI engines choose what to cite presented the 2023 Princeton gains (for example, up to 40% for expert quotes) as citation levers. They are now labeled as a 2023 benchmark that did not replicate on 2026 engines, and we recommend statistics and quotes for credibility only.
• A wrong credit. The same guide credited "44.2% of ChatGPT citations come from the first 30% of a page" to Zyppy. The source is Kevin Indig's analysis in Growth Memo, reported by Search Engine Land. The number stands; our credit was wrong.
• Multipliers we couldn't support. We removed "2.5x more citations" for content updated in the last 30 days from our AI platform playbook; a 30-day retrieval claim and "~3.2× more citations" for original data from our Perplexity guide; and schema citation lifts of 23% to 47% from the Perplexity guide and our AI citation gap guide. We found no published method behind the freshness figure, and Ahrefs' test found no significant lift from schema.
• Platform facts that changed. The playbook called Perplexity "the highest-opportunity platform for business websites"; with 0.9% of generative AI web traffic, it's worth the effort only if your customers use it. Our llms.txt guide said Google "has not confirmed support"; Google now says Search ignores llms.txt. Our measurement guide and AI crawler guide no longer list ChatGPT Atlas, which OpenAI announced it was shutting down on July 9, 2026 (TechCrunch).
• Measurement advice. The measurement guide called Bing's report the only free, official source of AI citation data; Search Console now reports Google's AI impressions too. We also told readers to build a custom GA4 segment for AI traffic, including in GEO vs SEO; GA4 now has a default AI Assistant channel.
• A dated conversion figure. The 4.4× AI-traffic conversion figure is now dated "Semrush, 2025" in the four guides that cite it. Semrush's June 2026 AI visibility index contains no conversion figures.

When the evidence moves again, we'll update this article and say what changed.

What held up this quarter isn't a secret technique: be findable, keep your own pages accurate, and measure honestly. If you want to know which of these foundations your business needs first, start with an assessment with a defined scope: Request my assessment.