SEO is becoming GEO: brands now fight to be cited, not just ranked

The ranking is no longer the result

For years, the job was simple to describe: rank high, get clicked. That chain is breaking. Writesonic explains that citation analysis now tracks which websites ChatGPT links to when it answers a tracked prompt, and the results are unstable: in its research, the top-ranked brand changed 52% of the time on the same prompt. A brand can win the answer on Monday and lose it on Tuesday without changing anything. The implication for creators is plain: a single good ranking is no longer an asset you hold, it is a position you keep re-earning. m:…4065

The same study gives a sense of how narrow the window is. Writesonic found that GPT-6 Astra, one of the newest ChatGPT models, ran an average of 6.1 fan-out questions per prompt and 41% fewer searches than the model before it, while reading 50 pages and citing only five sources. Most of what the model reads never gets credited. Being read is not the same as being cited, and being cited is what puts a brand in front of the reader. m:…4065

The plumbing decides who gets cited

Pixis argues the deciding work now sits in the backend, in what it calls technical GEO: crawlability, indexation, structured data, performance, and content structure. Its rule is blunt: a model cannot cite a page it cannot crawl, cannot parse a page it cannot render, and cannot trust a page whose structured data contradicts what is on screen. Pages that block the retrieval crawler never appear at all. Writesonic makes the same point about OAI-SearchBot: pages that block it cannot appear in ChatGPT search answers. m:…3481 m:…4065

One confusion Pixis flags is common and costly. Training crawlers such as GPTBot and ClaudeBot feed model training, but they do not affect AI search citations. Retrieval crawlers such as OAI-SearchBot, Claude-SearchBot, PerplexityBot, and Bingbot build the live answers, and those are the ones that decide citations. Pixis also recommends JSON-LD for structured data because it is easier to maintain than microdata, and says the schema has to match the visible page: stale prices or a rating that no longer applies read as a mismatch and lower trust rather than raising it. This is ongoing work, not a one-time setup, because crawl access, schema validity, and rendering drift over time. m:…3481

When the answer is unflattering

Being cited is only half the problem. Pixis sets out a five-step framework for auditing negative brand sentiment in AI answers, and its first distinction matters: negative sentiment and factual inaccuracy are separate findings, and they need different fixes. A wrong fact can be corrected with a supported update to owned content. Accurate criticism, Pixis says, may require a business change rather than a content correction. And even a clean correction carries no guarantee: updating owned website content gives retrieval systems current information to find, but it does not guarantee that an answer will use it. m:…3480

Where the audience actually is

Some of the most-cited material is not owned by brands at all. Pixis says AI engines such as ChatGPT and Perplexity frequently cite Reddit and Quora threads when answering product-related queries. Reddit was the top cited domain in an Ahrefs AI Mode dataset Pixis cites, accounting for 17.9% of citations, and Reddit reported 52.5 million US daily active uniques in the fourth quarter of 2025. But the ground shifts fast: Pixis notes that ChatGPT's citation of Reddit fell from 60% in August 2025 to 10% by mid-September 2025. Quora threads analyzed by Semrush averaged 37 answers, with answers at least 535 words long. Pixis advises brands to disclose affiliation in every answer, budget three to five hours a week per brand, and never coordinate upvotes, which violates platform rules and can lead to permanent bans. m:…1647

The measurement tools are changing too, and not in the direction marketers hoped. LLM Pulse reports that OpenAI is introducing a ChatGPT Business Hub that lets businesses claim their presence and view two daily metrics: impressions in ChatGPT answers and clicks to their website. The author argues that granular keyword-level data will not return, because AI search platforms are unlikely to provide prompt-level data given privacy risks and the lack of aggregation in user prompts. The same piece reports ChatGPT's share of worldwide visits to generative AI websites fell from 76% in June 2025 to 53% in May 2026, while Gemini reached about 27% and Claude about 9%. Its advice is to track share of voice across a fixed set of prompts instead of waiting for keywords to come back. m:…5020

What to do with this

The practical shape of GEO is now clear enough to act on. Keep retrieval crawlers able to reach the pages that matter, keep structured data in step with what the page actually shows, and lead with the direct answer before the supporting detail, which Pixis calls the most effective pattern for being cited. Check citations weekly on the prompts that matter, compare against Gemini or Perplexity when numbers drop to see whether the loss is yours alone, and treat third-party sources as part of the strategy: Writesonic says a healthy citation spread mixes owned pages with trusted third-party sources such as review sites or Reddit threads. The ranking still matters, but it is no longer the whole job. The job is the answer. m:…3481 m:…4065