# GTM digest: sources, AI and operating budget

> *AI-generated recommendations for discussion. These are not Nick’s recommendations or approved JLL/PRISM guidance.

Current configuration verified September 29, 2026. These are application limits and a usage calculation, not a provider invoice, current account balance or promise of free operation.

## What the digest does

After Monday's 13:00 UTC collection, the site generates a dated GTM digest with reported competitor moves, suggested PRISM actions, reasons, possible tests and market context. Read [the digest](https://crefieldnotes.com/gtm-digest), [Markdown](https://crefieldnotes.com/gtm-digest.md) or [JSON](https://crefieldnotes.com/api/gtm-digest). Older editions are retained.

The model is **Gemma 4 26B-A4B on Workers AI**. Input is bounded to up to three dated excerpts per discovery source, up to twelve deduplicated review-change observations, and latest available structured macro observations. Initial review baselines are excluded. Structure and citation-ID validation do not verify every article or recommendation. Missing or undated results do not establish competitor inactivity.

## Current application caps

| Provider or function | Calendar-month limit in code | What shares it |
|---|---:|---|
| Firecrawl | 900 attempts | Product/documentation scraping, directory reviews and Apple-feed fallbacks |
| Exa | 60 searches | Eleven enabled scheduled discovery sources plus manual recovery |
| Apify social collection | 25 task starts | Five configured weekly tasks; requested per-run maximum $0.30, with Cove X $0.10 |
| Apify Google Play reviews | 70 starts | Two daily listings; requested per-run maximum $0.10 |
| Social/review inference | $10 accounted or reserved | Gemini 3.1 Pro Preview through OpenRouter; both cohorts use the same allowance |
| GTM digest generation | 12 attempts | Scheduled and manual digest attempts; Workers AI billing is separate |

These controls are local to this application. Other projects can consume the same provider credits; actor pricing, account charges, storage and provider billing rules can differ. Unknown-cost inference failures retain their reservation. No subscription change was made as part of this documentation update.

## Will the limits cover a full month?

The current enabled registry routes 151 weekly sources to Firecrawl, 224 to public documentation APIs, two to public feeds, eleven to Exa, five to Apify social tasks and eight to review adapters.

- In a five-Monday, 31-day month, 151 weekly Firecrawl sources use **755 attempts**, and four daily G2/Software Advice listings add **124**, leaving **21 of the 900 attempts** before Apple fallbacks, new articles or manual recovery.
- If both Apple listings need both lookup and review-feed fallbacks daily, that can add **124 more attempts**. The same example becomes **1,003**, above the application cap. Coverage can therefore stop before the month ends under the current configuration.
- Eleven weekly Exa queries use up to **55 searches** in a five-run month; manual searches share the remaining five.
- Five weekly Apify social tasks use all **25 starts** in a five-run month. Two daily Google Play tasks use **62 of 70 starts** in a 31-day month.

Actual source growth and failures affect usage. Check provider dashboards and [live collection health](https://crefieldnotes.com/updates) before changing an allowance. Increasing the application cap alone does not purchase provider credits or make restricted pages accessible.

## Social and review analysis

**Gemini 3.1 Pro Preview via OpenRouter** analyzes the full available captured text within the configured size limit. It records topics, sentiment targets, attributed claims, exact source quotes and uncertainty. Exact-quote checks reject invented snippets; they do not prove an interpretation correct.

The system caches by content, model and prompt version. It allows up to ten uncached analysis attempts per company-post capture and five per review-source capture. Social and review analysis share the $10 monthly allowance. Oversized text is retained without silent shortening and can remain unanalyzed. Company-owned messaging is not customer satisfaction; directory, tenant-app and pilot-comment cohorts remain separate.

The model request limits provider prices to $2 per million input tokens and $12 per million output tokens. These are configured request ceilings, not a claim about current market pricing. Application accounting uses returned usage costs. [OpenRouter model page](https://openrouter.ai/google/gemini-3.1-pro-preview).

## Connected sources

| Source | Current coverage |
|---|---|
| Cove and Visitt documentation, product pages and release hubs | Detailed public-source monitoring; newly discovered official help articles are enrolled |
| Equiem, HqO, MRI Angus and Yardi | Selected public product/customer/AI pages and discovery, with less depth than Cove/Visitt |
| Company social accounts | Cove LinkedIn, X, Facebook and YouTube; Visitt LinkedIn and YouTube. Five Apify tasks plus Cove YouTube RSS |
| Customer reviews | Eight daily Cove/Visitt G2, Software Advice, Apple and Google Play listings. Capterra/GetApp syndicated pools are not separately polled |
| Official economic data | Four FRED series: DGS10, UNRATE, CPIAUCSL, DRCRELEXFACBS |
| Public CRE reports and news | Exa discovery includes official data and research domains such as JLL, CBRE, Cushman & Wakefield, NAIOP and Trepp; this is not a licensed comprehensive data feed |
| Audience comments | Historical capped pilot across selected platforms; no recurring collector |

FRED values retain observation dates, units and retrieval vintages. Monthly/quarterly data is not a new weekly release merely because it was fetched that week. National indicators do not measure a buyer's local rents, vacancy, absorption or ROI.

## Useful additions that are not connected

Potential additions include a scoped REIT/SEC earnings watchlist, licensed local CRE data, selected local building rules, dedicated podcast/webinar transcripts and recurring audience-comment collection. CRM win/loss evidence, customer calls, Teams and private PRISM code would require separate authorized internal access. These are possible extensions, not shipped integrations.

For a team discussion, connect the source fact to a buyer pressure, a proposed PRISM response and a small validation test. Keep proposed product advantages separate from observed competitor behavior. [Current coverage and limitations](https://crefieldnotes.com/LIVE-MONITORING.md).