- AI visibility KPIs already establishes the method that matters: a fixed prompt set, repeated enough times to see a stable percentage rather than a single check. This piece is the literal build: the real API options, real costs, and real limitations per platform.
- OpenAI, Anthropic, and Perplexity all have documented, usable APIs with citation data attached to web-search responses. Google does not have a public API for AI Overviews specifically. That's the one gap a DIY setup cannot close.
- Every platform with an API has a real, independently-reported gap between what the API returns and what the actual consumer chat product shows for the identical query. This isn't a theoretical caveat; it's been reported on OpenAI's and Perplexity's own developer forums, and stated directly in a competing vendor's own published methodology.
- At a realistic tracking cadence (12 prompts, 60 runs each, across four platforms, weekly), the actual cost comes out to roughly $50-200 a month total. Cost and rate limits are not the real obstacle. Engineering time to build and maintain the pipeline is.
- Several real, current open-source projects already do versions of this, ranging from API-based trackers to ones that deliberately reject APIs in favor of browser automation specifically to capture what a logged-in user actually sees.
AI visibility KPIs already lays out the method: a fixed set of real customer-language prompts, run across the platforms that matter, repeated enough times per prompt to see a stable percentage instead of a single yes/no check, the same discipline SparkToro and Gumshoe.ai's research showed is necessary, since a single prompt run has under 1-in-100 odds of matching a second run exactly. This piece is the literal build: what it actually takes to run that methodology yourself, platform by platform, with real costs and real limitations.
What are you actually building?
The mechanics are the same regardless of platform: send a fixed prompt to an API, capture the response text and any citation data it returns, log it with a timestamp, and repeat on a schedule. Over enough runs, you get the appearance-frequency signal the KPI post describes: how often your brand shows up, not whether it showed up once.
What's actually available, platform by platform?
OpenAI (ChatGPT). The plain Chat Completions API has no live web access and returns nothing based on current information; it answers from training data alone. The relevant tool is the Responses API with the web_search capability enabled, which returns inline citation annotations plus a separate field listing every URL actually consulted. Pricing runs around $10 per 1,000 web-search calls, plus normal token costs on top. The real caveat: multiple developers have reported on OpenAI's own community forum that the API's web-search tool returns materially different, sometimes worse, results than ChatGPT's consumer search product for the identical query, including broken links the consumer product doesn't return. OpenAI hasn't published anything confirming whether the two share the same underlying retrieval index.
Anthropic (Claude). Claude's web search tool is documented clearly and directly: citations are automatic and always attached whenever the tool is invoked, at $10 per 1,000 searches plus standard token costs on top. Rate limits are set per organization and visible in each account's console; a large batch job run all at once may need pacing rather than firing everything simultaneously. Whether claude.ai (the consumer product) uses the identical retrieval mechanism as the API isn't explicitly confirmed by Anthropic either way.
Perplexity. The Sonar API family (sonar, sonar-pro, sonar-reasoning-pro, and others) returns citation data as a standard field at no separate per-search fee, cheaper structurally than OpenAI's or Anthropic's per-search pricing, just ordinary token cost. The same consumer-versus-API gap shows up here too: a user reported on Perplexity's own community forum getting a confidently wrong answer from the API on a query where the consumer web interface correctly hedged, for the identical prompt. Separately, Perplexity's citation ordering can shift between identical runs, and the raw API response strips out UI elements (related-question cards, shopping cards) that are part of what actually drives visibility in the real product.
Google (Gemini). This is the one genuine, unclosable gap. Gemini's API offers "Grounding with Google Search" as a tool, returning source metadata when the model decides to search. Gemini 3.x models include 5,000 free grounded prompts a month before billing kicks in at roughly $14 per 1,000 after that. But there is no public API for Google AI Overviews itself. AI Overviews are a feature built into Google Search, not a callable product surface on their own. Gemini's Grounding tool draws on the same underlying Google Search index by Google's own description, but it's a different product, and using it in anything shown to end users carries an obligation to display Google's "Search Suggestions" UI element, a real constraint even for an internal dashboard. If your platform mix needs to include AI Overviews specifically, there's no honest way to fully replicate that exact surface through any documented API. You'd be measuring Gemini-the-model, not AI-Overviews-the-search-feature, and a build should say so plainly rather than presenting the two as interchangeable.
What's the cost math for a real tracking cadence?
Twelve prompts, run 60 times each, across four platforms, weekly, works out to 720 queries per platform per week, roughly 3,120 per platform per month, about 12,480 total across all four. At current pricing: OpenAI comes to roughly $40-100 a month including search fees and token cost; Anthropic lands in a similar $40-90 range; Perplexity, with no separate search fee, is likely the cheapest of the four at well under $20; Gemini can land at effectively $0 if the volume stays under its free monthly grounded-prompt allotment, or a modest amount once past it. Total, realistically: somewhere in the $50-200 a month range across all four platforms, well inside reach for a solo operator or small team. Rate limits aren't the binding constraint either: even the most restrictive free tier found comfortably covers this cadence if calls are paced with a few seconds between requests rather than fired in a burst. The actual obstacle is the engineering time to build the runner, normalize each platform's different response format, and extract brand mentions consistently, not cost or throughput.
What can't this setup genuinely do?
Session and personalization effects are a real, separately-documented limitation. Gumshoe.ai (the company behind the SparkToro consistency study already cited on this site) states directly in its own published methodology that it uses official APIs specifically because that's the only way to simulate logged-in personalization at scale in a compliant way, and separately notes the consumer web interface applies system prompts that shape response length, formatting, and citation behavior differently than a raw API call without them. That's a competing vendor's own stated limitation, not speculation, and it applies to a DIY build exactly as much as it applies to a paid tool built the same way.
What if you'd rather not build from scratch?
Several current, real open-source projects already attempt versions of this, with genuinely different philosophies worth knowing about before picking one:
- Tools built around official APIs (OpenAI, Anthropic, and routing layers like OpenRouter that call multiple models through one interface) get you the mechanism described above with less setup work, at the cost of the same API-versus-consumer-product gap.
- At least one actively-maintained project takes the opposite approach on principle: driving a real, logged-in browser and reading the rendered page instead of calling any API, specifically because its authors argue an API or headless read "measures a surface nobody sees." That's a direct, practitioner-level acknowledgment of the same gap described above, solved by trading API convenience for browser-automation complexity.
Neither approach is strictly better. API-based tools are faster to build and cheaper to run; browser-automation tools capture what an actual logged-in user sees, including the UI elements a raw API strips out, at the cost of being slower and more fragile to maintain.
Frequently asked questions
Can I really track AI visibility without paying for Profound, Peec AI, or Otterly?
Yes, for three of the four major platforms. The API access, cost, and rate limits are all within reach of a solo operator or small team, likely $50-200/month total. The real obstacles are engineering time and one specific, unavoidable gap: there's no public API for Google AI Overviews, so you can't fully replicate that one surface this way.
Do the official APIs return the same results as the consumer chat apps?
Not reliably, and this is documented in multiple places, not just assumed. Developers have reported on OpenAI's and Perplexity's own community forums that API responses differ from what the consumer product shows for an identical query. Even Gumshoe.ai (the company behind the SparkToro consistency study this site already cites) states in its own published methodology that raw API calls without the consumer app's system prompt produce different response length, formatting, and citation behavior.
What can't a DIY setup track at all?
Google AI Overviews specifically. There is no public API for that feature: it's a capability embedded inside Google Search itself, not a callable product. The closest available substitute, Gemini's Grounding with Google Search tool, uses the same underlying search index by Google's own description, but it's a genuinely different product surface, not AI Overviews itself, and using it also carries a UI-display obligation worth knowing before you build around it.
How much does it actually cost to run this yourself?
For a cadence like 12 prompts run 60 times per prompt across four platforms, weekly, the arithmetic works out to roughly 3,000+ queries per platform per month. At current per-platform pricing that's realistically $50-200 a month total across all four APIs, well within reach for a small team. The binding constraint is the engineering time to build and maintain it, not the cost or the rate limits.
Sources referenced
- Anthropic: Web search tool documentation
- SparkToro / Gumshoe.ai AI recommendation consistency study (Jan 2026, ~2,961 prompt runs)
This piece is the practical build behind the measurement methodology in AI visibility KPIs. For the theory of what's stable enough to track in the first place, start there.
