What we're building, how it works, and how each AI engine actually decides who it recommends. We keep adding to this as we ship.
BestSearch is an AI visibility intelligence platform. As people research using Perplexity, ChatGPT, and Gemini instead of Google, brands need a layer beyond traditional SEO to know:
Each entity gets a weight by rank: rank 1 = 10 pts, rank 10 = 1 pt.
Σ rank_weight + per-engine bonus (Perplexity +2, OpenAI/Gemini +1).
mentions × engines_count × avg(rank_weight) — rewards entities every engine agrees on.
Top-3 rank variance across the last 5 scans of the same query → Stable / Moderate / High volatility.
Perplexity runs a live web search for each query, reads the top results, and summarises them. If you're not in the pages it retrieves, you don't exist to it.
→ Get listed on the domains Perplexity cites — Sources tracks exactly which ones, per query.
GPT answers from memory baked in during training — it has no internet access in this mode. Brands appear because they were mentioned repeatedly in Wikipedia, news archives, and high-authority content before the training cutoff.
→ Long-term authority content (Wikipedia, press, comparison articles) moves this engine, not this week's blog post.
Same GPT-4o mini model, but with OpenAI's Responses API web-search tool enabled — effectively ChatGPT's browse mode. It searches live and returns real citation URLs, like Perplexity.
→ Behaves like a search engine crawl target, not a training-data target — recent, indexable pages matter here.
Gemini grounds its answers in Google's knowledge graph and, for some queries, real-time Search. Brands show up because they have a verified Google Business Profile, a Wikidata entity, and structured markup on their site.
→ Structured data (schema.org, Business Profile, Wikidata) is the highest-leverage fix for this engine.
Claude answers from parametric memory built during training. It favours authoritative, factual web content — Wikipedia, comparison articles, and curated long-form pieces on high-authority sites. No live web access in standard mode.
→ Same lever as OpenAI's base model: durable, citation-worthy content on sites Claude's training data trusts.
AI vendors send crawlers to your site for different reasons. LLM Analytics classifies every hit into one of three categories:
Crawl content for potential use in training AI models (e.g. GPTBot, ClaudeBot).
Index your site so an AI can cite it in a live search answer (e.g. OAI-SearchBot, PerplexityBot).
Visit a specific page only because a real user's prompt required it — a live citation signal (e.g. ChatGPT-User, Perplexity-User).
Full reference table with every known bot is in LLM Analytics.
Your buyers are already discussing your category on Reddit. Reddit Marketing surfaces those threads, scores which subreddits are worth your time, and drafts a reply for you to review — it never posts anything on its own.
There is no submit/comment/vote scope anywhere in this feature — lib/reddit/client.ts only ever calls Reddit's GET endpoints. Reply drafts are exactly that: drafts. Undisclosed self-promotion is against Reddit's rules and is how "marketing tools" turn into astroturfing bots — the prompt is written to avoid that by default.
Building this in the open — full architecture, the fit-score heuristic, environment variables, and known limitations are tracked in md/reddit-marketing/engine.md, updated as it grows. Try it at Reddit Marketing.
This log grows as we ship — check back for what's new.