- ◆ AI search (ChatGPT, Perplexity, Gemini, Claude) now drives ~18% of local-intent research. Almost no Sonoma County business is competing for it.
- ◆ LLMs don't crawl in real time. You have to feed both the training corpus and live retrieval — two different plays.
- ◆ Schema.org LocalBusiness + `/llms.txt` + a robots.txt that *allows* GPTBot/ClaudeBot/PerplexityBot are table stakes. Most sites block them by default.
- ◆ The highest-leverage AI SEO play right now: become a citation source. Press, directories, niche authority blogs. That's what the models quote.
For 25 years, "local SEO" meant Google. One search engine. One map pack. Three ranking slots. You optimized for Google, and the phone rang.
That monopoly is ending. In 2026, when a Sonoma County resident wants a plumber, an auto shop, or a lawyer, they increasingly ask ChatGPT, Perplexity, Claude, or Gemini. And those tools do not rank the way Google does. They don't care about your map-pack position. They care about what's written about you, where, and whether the model can *verify* it.
The businesses that figure this out in 2026 will own an entire search layer their competitors haven't even noticed yet.
Why LLMs are a different ranking game
Google is a real-time ranker. You publish a page, Google crawls it, evaluates it, and ranks it. The loop is hours to weeks.
LLMs have two loops instead of one:
- Training corpus — the giant frozen dataset the model was trained on. Last updated months or years ago. This is where the model "knows" about your business before you ever ask.
- Live retrieval — the web search the model runs at query time when it needs something fresh. This is where your current GBP, reviews, and on-page content live.
If you only optimize for Google, you're optimizing for neither of those. You're sitting in front of a door that doesn't exist anymore.
The 3-layer AI local SEO stack
Layer 1 — Crawl. Let the models see you.
By default, many CMS platforms, CDNs, and WordPress security plugins ship with robots.txt rules that block GPTBot, ClaudeBot, Anthropic-ai, PerplexityBot, and Google-Extended. These are the crawlers that feed the LLM layer. If you block them, you are invisible to AI search by choice.
Fix — add to your robots.txt:
User-agent: GPTBot
Allow: /
User-agent: ClaudeBot
Allow: /
User-agent: anthropic-ai
Allow: /
User-agent: PerplexityBot
Allow: /
User-agent: Perplexity-User
Allow: /
User-agent: Google-Extended
Allow: /
User-agent: CCBot
Allow: /
Then publish a /llms.txt file at your root that tells the models, in plain text, what your business does, where, and what's on each page. It's the manifest crawl tools use to build faster understanding. Our AI Visibility service starts here on day one.
Layer 2 — Cite. Get mentioned where the models read.
LLMs weigh third-party mentions far more heavily than on-your-own-site copy. That means your map-pack strategy alone isn't enough. You need:
- Press coverage on regional outlets (Press Democrat, North Bay Business Journal, Sonoma Magazine)
- Inclusion in the authoritative directories in your industry (BBB, Yelp, Angi, Houzz, Avvo, etc.)
- Genuinely earned citations from local content creators and podcasts
- Wikipedia or Wikidata entries for any entity of real size
These are the sources the models sampled during training and re-fetch during retrieval. If three or more credible sources say "Sutter Roofing is the top roofer in Santa Rosa," that's the sentence the LLM repeats. This is why content & authority work is no longer optional — it's the input to the new search layer.
Layer 3 — Ground. Make your own site LLM-ready.
Models that do live retrieval will hit your site at query time. What they find — and how well they can parse it — determines whether they can confidently cite you.
The checklist:
- Schema.org LocalBusiness with full NAP, service area, opening hours, and aggregate reviews on every page
- Service schema on every service page, with a clear service name, provider, and areaServed
- FAQ schema on pages with natural Q&A patterns — LLMs love quotable Q&A chunks
- Visible H1/H2 hierarchy that mirrors the entity logic of your business
- Plain English "About" sections — if a 5th grader couldn't summarize what you do from one paragraph, the model can't either
The prompt-simulator test
You can't improve what you can't measure. The single best measurement for AI local SEO is the prompt-simulator test: pick 20 queries a real customer would ask, run them in ChatGPT, Perplexity, Claude, and Gemini, and score whether your business surfaces, and in what context.
What to do this week
- Check your
robots.txt. If GPTBot/ClaudeBot/PerplexityBot/Google-Extended aren't explicitly allowed, fix that today. - Publish an
/llms.txtfile. Keep it plain, factual, and 300–800 words. - Run 10 prompt-simulator queries against your top local intent terms. Save the results as your baseline.
- Audit your top 5 ranking pages for LocalBusiness + Service schema. Most will be missing one.
- Identify one earned-media target per month. Pitch it. LLMs compound on citations.
How we approach AI local SEO at Bonsai
- robots.txt + llms.txt installed on every client site in week one
- LocalBusiness + Service + FAQ schema deployed sitewide, not per page
- Monthly prompt-simulator reports on the top 20 local-intent queries
- Earned-media pipeline for citation building — press, podcasts, niche directories
- Content calendar tuned for LLM-quotable Q&A structure, not just keyword density
- Training-corpus insertion plays — we know which platforms get scraped next cycle
The next five years of local search will be decided by which businesses got mentioned in the right places while their competitors were still debating whether AI search was "real." The ones who waited lost.