Data to AI™ / Gemini/Facebook Ads
Pull Meta ad spend, placement and audience figures into your terminal or VS Code while you build notebooks, scripts and dashboards, without dropping another Ads Manager download into the project.
Better answers, fewer tokens.
Gemini CLI
Which placements returned the most purchase value per dollar spent last month?
Instagram Reels returned 4.1 in purchase value per unit of spend, the best of the three, on about a third of the budget. Facebook Feed took the largest share at 3.2.
| Placement | Spend | Purchase Roas | Ctr |
|---|---|---|---|
| Instagram Reels Platform Position Device Ad Insights | $8,400 | 4.1 | 1.32% |
| Facebook Feed Platform Position Device Ad Insights | $12,600 | 3.2 | 1.05% |
| Instagram Stories Platform Position Device Ad Insights | $5,200 | 2.4 | 0.71% |
Spend and Purchase Roas come from one placement breakdown the Facebook Ads API supports, not from separate exports matched by hand.
Facebook Ads extracts paid performance from Meta ad accounts, from account down to ad level: spend, delivery, conversions and ROAS, broken down by placement, device, demographics or creative asset.
Query Age Gender Breakdowns Ad Insights for September and save Age, Gender, Spend and Purchase Roas to data/meta-audiences.csv.
A marketing analyst starts Gemini CLI inside the notebook project, approves the query call and loads the file straight into pandas. Next month the same prompt rebuilds the CSV from the scheduled flow, so the charts never depend on whoever exported last.
Our loader in src/metrics/meta.ts divides Purchase (value) by Spend. Pull last week's Facebook Ads Insights per Campaign name and tell me where our ROAS differs.
In Gemini Code Assist agent mode, a developer compares the dashboard's output with figures read from Dataddo before the pull request is merged. Mismatches show up in review, not in the Monday meeting with the paid social team.
Read Ad Name, Creative Title, Body and Creative CTA Type from Ad Creatives (by Ads) and write scripts/creative_audit.py that flags duplicate titles.
A technical marketer gets a small, rerunnable script instead of a one-off answer. It uses the exact field names and runs again after every new batch of ads.
Data to AI™ is priced per seat, like the AI assistants your team already uses, with ~25% off on annual billing. Need tailored contracts or volume pricing? Talk to us about Enterprise.
Serve governed business data to Claude, ChatGPT and other AI tools over MCP.
$10.49/month
per seat
3 seats minimum
Flexibly deploy Dataddo in any cloud or hybrid environment.
Custom
We have a payment model that works for you
| Without Dataddo |
With Dataddo
|
Outcome for you | |
|---|---|---|---|
| Meta data in a project folder | Someone downloads a CSV from Ads Manager, copies it into the repository and renames columns before any script can read it. | Gemini CLI asks the dataddo server for Facebook Ads datasets by name, so code starts from fields such as Spend and Purchase Roas. | No hand-copied files |
| Which projects can reach ad data | A shared export ends up in every repository, whether the project needs Meta numbers or not. | Register the server in .gemini/settings.json of a single project, or in ~/.gemini/settings.json when every project should see it. | Access scoped per project |
| Watching each lookup | A script with stored Marketing API credentials runs whatever it was coded to do, and nobody sees the calls. | Gemini CLI asks you to confirm every Dataddo tool call, and Always allow is chosen per tool, for example only for query. | Every call is visible |
| Terminal and editor | The terminal session and the IDE each work from a different copy of the ad data, exported on different days. | Gemini Code Assist agent mode in VS Code reads the same settings.json, so the editor and Gemini CLI query identical Facebook Ads flows. | One setup for both tools |
| Token consumption | Gemini spends tokens exploring raw Facebook Ads columns, sampling rows and retrying until it understands the data. | Field definitions and relations are supplied up front, so the model skips the exploration and goes straight to the answer. | Fewer tokens per answer |
| Answer quality | Prone to wrong joins, invented fields and numbers that are hard to check. | Answers grounded in defined fields. In the Dataddo benchmark, 78.5% of questions were answered correctly, against 65.5% with plain CSV files. | Better answers |
The Facebook Ads datasets marketing teams use most, with their real field names. Gemini queries them by name through the Dataddo semantic layer.
| Dataset | Metrics and attributes |
Ask Gemini
|
|---|---|---|
| Facebook Ads Insights |
|
Write Spend, Impressions and CTR per Ad Set name for the last 14 days to a CSV I can load into a notebook. |
| Platform Position Device Ad Insights |
|
Which Publisher Platform and Platform Position pairs had Spend last week but a Purchase Roas below 1? |
| Age Gender Breakdowns Ad Insights |
|
List the Age and Gender rows with the highest Reach last month as a CSV for our audience notebook. |
| Video Asset Ad Insights |
|
Which Video Asset rows had the most Video Thruplay Watched Actions last month, as a markdown table? |
| Ad Creatives (by Ads) |
|
Export Ad Name, Creative Title and Creative CTA Type for every ad to a JSON file for our copy checks. |
| Campaigns Insights |
|
Which campaigns have a Daily Budget above the average for campaigns with the same Objective? |
We asked 43 pre-registered questions of three real datasets - Google Search Console, HubSpot deals and Google Ads - and delivered the same rows to the same model in four ways. Only the data layer changed.
of questions answered correctly - the best of the four delivery methods tested
fewer confidently wrong answers - 14.0% of runs against 24.0% and 26.3%
more correct answers on messy CRM data - better on 8 questions, worse on none
Same model, prompts and rows in every condition; 759 graded runs, ground truth frozen before any run. On Google Search Console alone, documented CSV files did as well as Data to AI™ - the advantage comes from messier data such as HubSpot deals.
Read the benchmark →In Dataddo, create a Google Gemini authorizer, a Facebook Ads source with datasets such as Facebook Ads Insights, and a Google Gemini destination or AI Model. Add Dataddo under mcpServers in ~/.gemini/settings.json with httpUrl https://headless.dataddo.com/mcp-data, restart Gemini CLI and run /mcp auth dataddo to sign in.
Check the key name. The entry must use httpUrl, because url makes Gemini CLI fall back to SSE, which Dataddo does not support. Leave out headers and any oauth block, restart Gemini CLI, then run /mcp auth dataddo.
No. The Dataddo tools only read Facebook Ads data already extracted into your flows, through the semantic layer. Dataddo never accepts SQL from the model, and budgets, bids and ad statuses in Meta stay exactly as they are.
No. The sign-in page asks for your Dataddo email and password, with MFA if your account uses it; Google and Microsoft SSO are not offered here. Gemini CLI opens that page in a browser, so run it on a machine that has one.
Yes. Once signed in, Gemini reaches every flow or Context attached to an AI destination or AI Model in that Dataddo account, not only the Google Gemini ones. Attach to AI destinations only the ad accounts you are happy for it to query.
Gemini Code Assist in IntelliJ uses its own mcp.json instead of settings.json. Add the same Dataddo server there; in VS Code, agent mode picks up the settings.json you already use for Gemini CLI.
Tokens live in ~/.gemini/mcp-oauth-tokens.json and renew automatically. If a session expires or is revoked, run /mcp auth dataddo again, then check /mcp to confirm the dataddo server is connected.
