AI Providers and Models

AI Providers and Models

Every AI feature in OpenGSC runs on your own API key, at your provider's price, with no markup. Keys live in Settings → AI models.

Supported providers

Provider	Notes
Anthropic (Claude)	Also used by AI Visibility (Claude engine)
OpenAI	Also used by AI Visibility (ChatGPT) and GEO Audit
Google Gemini	Also used by AI Visibility (Gemini) and GEO Audit
OpenRouter	Many models behind one key
Cheaper Inference	Low-cost relay. Billing errors can come back as 502
Z.AI (GLM)
Kimi (Moonshot AI)
Kie.ai	Also a GEO Audit engine
DeepSeek
Qwen (Alibaba Cloud)
Custom endpoint	Any OpenAI-compatible API: base URL + key + model

Each provider card fetches the live model list from the provider's API, so you choose from what your account can actually use.

Perplexity and xAI (Grok) keys are only used for AI Visibility and live in Settings → AEO & site health.

Per-task models

Every AI job is a named task. Each task can use its own provider and model, or inherit the suite-wide default.

Task	What it does	Default intent
Structure (outline)	Drafts H1/H2/H3 and what goes under each	best available
Article text	Body copy: articles, rewrite, humanize	best available
Landing page	Commercial page copy and wireframes	best available
Content analysis	Reads pages and SERPs, extracts findings	everyday
Editorial policy	Drafts policies from your brand pages	everyday
QA judge	Independent check that a finished article/outline is complete and publishable. Best set to a different model than the writer, because a same-model judge shares the writer's blind spots	everyday
Utility passes	Short mechanical passes whose output is parsed (e.g. GEO Audit's structured second pass)	cheap
Domain history AI	The Drops AI history verdict	cheap
Publish respin	Adapts a post for one platform before publishing	cheap

Defaults aren't hard-coded model ids. They're intentions ("best available / everyday / cheap") resolved against your account's current model list, so they don't go stale when a provider ships a new generation.

Which model runs? The header of every SEO tool names the model that will run. Expand it to see each task the page performs, what it resolved to, and which settings level the value came from (task override, suite default, provider default). If an override seems to be ignored, check there first.

Other settings on the same tab

• Default model for the suite, and the custom OpenAI-compatible endpoint card.
• Fact-check: auto fact-check when a text is opened (uses SERP quota), only for sections with facts, sources scraped per section, reuse of the competitor corpus.
• Strip invisible AI marks (on by default): zero-width spaces, bidi controls, soft hyphens, exotic spaces.
• Page-mechanics check: a deterministic check after writing (mixed-alphabet words, prices in the wrong market's currency…), with an optional one-call repair.
• Hard-redact banned words: replaces policy banned words with […] for critical commercial pages.
• Auto image prompts when an article is opened (one cheap call).

Costs and safety

• Only Admin and Owner roles can run paid AI actions.
• Every outbound call is recorded in Settings → Provider log with status and, where the provider reports it, cost.
• The MCP tools that call AI (start_generation_job, start_rewrite_job) refuse to run without confirm: true. For most tasks a connected agent can do the writing itself for free; see MCP Server → Optimizing a page.

When generation fails

Look in History. The job's error carries the provider's real reason: invalid key, exhausted quota, context length on a huge outline, or the provider's content-policy filter rejecting the topic. pm2 logs opengsc has the same information on lines starting with [LLM].