--- title: "The Hidden Language of Search" description: "AI answer engines rewrite your prompts into queries. Understanding this translation layer explains the weird keywords in your GSC." tldr: "There's a hidden layer between human questions and search results. AI tools translate messy prompts into precise queries - and you can see the evidence in Google Search Console." date: 2026-03-02 tags: [AI, SEO, SEARCH, TECH] draft: false author: "Nikola Balić" topics: [AI Search, Google Search Console, RAG, Query Rewriting] entities: [ChatGPT Search, Perplexity, Google Search Console, OpenAI] answers_questions: - How do AI answer engines like ChatGPT Search retrieve information? - Why am I seeing weird long-tail keywords in Google Search Console? - What's the connection between AI prompt rewriting and SEO? --- A search query showed up in Google Search Console recently: > "browser-use open source agentic ai framework github repository technical documentation showing dependencies, foundation models supported, playwright integration, python libraries, and implementation architecture" Thirty-one words. No human typed that into Google. I’ve been using our new Steel CLI and `steel-browser` skill to explore this kind of case in practice. This demo shows Claude Code running **parallel browser sessions** with ChatGPT so you can inspect how it reasons and what answers it returns en masse. That string is a *translated* search query: the output of an AI rewriting someone's prompt into something a search engine can understand. And it's showing up in GSC because somewhere, an AI answer engine sent that exact string to Google. ## Two languages, one problem When you ask an AI tool a question, there are *two different languages* involved: 1. **Human language**: your prompt (messy, contextual, conversational) 2. **Retrieval language**: search queries (short, explicit, keyword-heavy) Most AI answer engines solve this by inserting a translation layer: **Prompt → (rewrite into query/queries) → Search → (select evidence) → Answer** OpenAI explicitly confirms this for ChatGPT Search: it "typically rewrites your query into one or more targeted queries" and may do follow-up queries after seeing initial results. This is documented behavior, not speculation. ### Why rewrite at all? Because raw prompts are terrible search queries: - "latest" needs a date/recency hint - "near me" needs location - vague nouns need disambiguation - multi-part questions need multiple searches This is well-studied in RAG research: rewriting, decomposition, and disambiguation improve retrieval quality. ## What the translation looks like ### Step 1: Interpret intent The AI first decides: *Do I need the web, or can I answer from training data?* ChatGPT Search automatically searches when your question benefits from web info. Perplexity is "search-first" by default. ### Step 2: Rewrite into queries ("fan-out") One prompt becomes one or more search queries. **Example from OpenAI's docs:** User: "what's the latest on drugs that target CCR8 for cancer?" Rewritten: "CCR8 immunotherapy drug development 2025" → then narrower follow-ups. **Another example:** User: "good restaurants near me" Rewritten with location: "top restaurants San Francisco" If ChatGPT Memory is enabled, it might add remembered preferences: "good vegan restaurants San Francisco." ### Step 3: Apply filters Some systems add constraints: domain, region, language. Perplexity's API exposes these controls explicitly. ### Step 4: Retrieve, dedupe, rerank The system merges results from multiple queries, removes duplicates, reranks by relevance/authority/recency, and opens pages to extract evidence. If evidence is missing? It iterates with another rewrite. ### Step 5: Synthesize with citations Finally, it writes a natural-language response grounded in what it retrieved. ## The evidence in your GSC Now the SEO-relevant part. That 31-word query I showed you? It has clear signatures of AI origin: - **Tool/code-like vocabulary**: "github repository", "implementation architecture" - **Long structured text**: 31 words, comma-separated clauses - **Multi-line/quoted snippet style**: reads like pasted context - **Connector tokens**: "showing", "and" chaining multiple requirements This is an AI *fan-out* query, the kind ChatGPT Search generates when someone asks a multi-part question about browser-use. And it's not alone. Here are more examples from real GSC data: | Query Pattern | Why It's Likely AI-Generated | |---------------|------------------------------| | "read https://better-auth.com/docs/concepts/rate-limit.mdx, i want to ask questions about it" | Contains URL + intent statement, not search syntax | | "anthropic claude computer use beta documentation" | Keyword-stuffed product name, no natural phrasing | | "playwright connect_over_cdp documentation python" | Underscore method name + language, very specific | | "which headless browser api should i integrate if i want an http endpoint my bots and llm agents can call on demand?" | Full question as query, 24 words | These queries have **zero clicks** but **impressions**. Why? Because they're so specific, they match few pages, but when they do match, your page shows up. ## Why this matters for SEO There are three practical implications here. ### 1. New keyword patterns are emerging AI-generated queries are: - Longer (20-40 words) - More structured (comma-separated, semi-colon delimited) - More specific (exact method names, versions, documentation paths) - Question-shaped but keyword-dense If you're seeing these in GSC, it's not spam. It's a new kind of traffic source. ### 2. Content should match AI query patterns Traditional SEO advice: write for humans, use natural language. New advice: *also* include the structured, keyword-dense phrasing that AI rewriters generate. Concrete tactics: - Add explicit query-style headers: "What is Steel?", "Steel vs Browserbase comparison" - Include technical specifics in headings: "playwright connect_over_cdp python documentation" - Create cluster pages that answer multi-part intents in one URL - Add temporal cues: "2026 benchmark", "March 2026 update" ### 3. Weird keywords aren't always weird Before you dismiss strange queries as noise, check: - Does it match your content technically? (method names, API endpoints) - Is it structured like an AI rewrite? (long, comma-separated, specific) - Does it have zero clicks but impressions? (high specificity = low volume) If yes, it might be AI-driven traffic, and worth optimizing for. ## The other explanation: security issues Not all weird queries are AI-generated. Some are warning signs. If you're seeing porn, pharma, or streaming keywords that have *nothing* to do with your site, check for: 1. **Hacked content**: page injection, content injection, cloaking 2. **Spammy URLs**: infinite parameter variants returning 200/OK 3. **The Japanese keyword hack**: auto-generated spam pages in random directories Google documents these patterns explicitly. They're real, and they show up in GSC as unrelated queries. The difference: AI queries are *topically relevant* but weirdly structured. Spam queries are *topically irrelevant* entirely. ## The bigger picture AI answer engines aren't replacing search. They're becoming a translation layer on top of it. When you ask ChatGPT a question, it doesn't just "know" the answer. It: 1. Rewrites your question into search queries 2. Sends those queries to search providers (including Google) 3. Reads the results 4. Synthesizes an answer Your content can appear in step 2, even if the human never visited Google directly. This is the new SEO frontier: **optimizing for AI rewriters**, not just human searchers. The evidence is already in your GSC. You just have to know what you're looking at. ## See this live If this is still abstract, watch the same workflow with the Steel CLI and `steel-browser` skill in the loop: You can also read the release context in [Steel CLI and the new Steel Browser skill](https://steel.dev/blog/steel-cli-and-agent-skill). --- ## Sources - [ChatGPT Search - OpenAI Help Center](https://help.openai.com/en/articles/9237897-chatgpt-search) - [Learning to Refine Queries for RAG - arXiv](https://arxiv.org/pdf/2404.00610) - [Perplexity Search API Documentation](https://docs.perplexity.ai/docs/search/quickstart) - [Question Decomposition for RAG - arXiv](https://arxiv.org/html/2507.00355v1) - [GSC Performance Report - Google Support](https://support.google.com/webmasters/answer/7576553) - [GSC Data Filtering Deep Dive - Google for Developers](https://developers.google.com/search/blog/2022/10/performance-data-deep-dive) - [Spam Policies - Google for Developers](https://developers.google.com/search/docs/essentials/spam-policies)