Why Your AI Marketing Agents Are Burning Tokens (And How to Fix It)

May 18, 2026

Summarize with AI:

I cut my AI marketing agent's token usage by 95%.

I was thinking of building a marketing agent on Claude to automate a repeating task. Instead, I used the Airbyte Agents UI to build the same agent, and the results were even better than what I thought was possible.

Running this same task directly through Claude Opus 4.7 would cost 300K-700K input tokens per run. On Airbyte Agents, it's 15K. That’s because the data stays in the Context Store instead of getting stuffed into the LLM's context window.

What I built is a paid search optimization agent that pulls our top Google Search Console queries, matches them against our live Google Ads keywords, analyzes the gaps, suggests the missing keywords and publishes a prioritized report to Notion on the 1st of every month.

Let’s look at how I built it and what the numbers actually look like. Also, this is not about just saving tokens but also building it very fast as the response time of Airbye agents is much faster because the data stays in the context store.

The problem I was solving

If you run paid search campaigns, you know there is always a keywords gap and you need to update the campaign on a weekly or monthly basis to keep it fresh. Think of stuff like queries where you are getting decent organic impressions on page two, but you are not leveraging them in your paid search campaigns. Commercial, high-intent queries are quietly showing up in GSC but completely missing from your Ads keyword list. Money sitting there wasted.

I used to do this analysis manually, maybe once a quarter because it’s painful and slow. Export GSC and then export Ads keywords and then match in a spreadsheet, then manually add everything to Google Ads. Once a quarter is definitely too slow, and every week or so that you’re not doing this, you’re wasting optimization potential that sits there till it’s too late.

What the automation does

The automation is called ‘Monthly GSC × Google Ads Keyword Gap Report.’ Runs at 8:00 AM PT on the 1st of every month and add this report on Notion:

Pull top 500 non-branded GSC queries, filtered to USA, web, for the last 30 days.
Then check all active Google Ads keywords and search terms to build the ‘covered’ keyword set.
Match GSC queries against live keywords on ads campaigns using exact/phrase/broad/search-term logic.
Create a score for each keyword 0–100 based on impressions, clicks and intent.
Pick the top recommended additions with suggested match type and campaign/ad group placement.
Check the quality score and make sure it's high or more than average.
Publish the full analysis report on Notion.

Here's the automation in the dashboard, scheduled and active:

And the full setup inside the Automation Builder. Three connectors (Google Search Console, Google Ads, Notion) and the cron schedule on the right:

All I did was describe what I wanted, set the trigger to monthly, and it built the 8-step workflow, the scoring logic, the error handling, and the Notion publishing. I didn't write anything about that except the prompt. Just make sure you connect to the required connectors on the platform.

The token math

This is what I was not expecting. I asked Claude to break down the token cost of running this job directly on Claude vs. on Airbyte Agents. Here’s what it told me:

The big rows are the data rows. On a direct Claude call, all your GSC queries, all your Google Ads keywords, all your search terms have to pass through the context window. 40K-60K tokens just for 500 GSC queries. 80K-200K for Ads keywords. 150K-400K for the search terms report.

On Airbyte Agents, that data sits in the Context Store. The agent queries the store, gets back small retrieved slices, and only those hit the LLM. GSC pull? ~0 tokens. Ads keywords? ~0 tokens. Search terms? ~0 tokens. All staying in the Context Store.

300K-700K tokens direct vs. 13K-15K on Agents. Around 20-50x reduction. And the gap gets wider as your keyword inventory grows, because direct scales linearly with data volume while Context Store stays roughly flat.

Here's what that looks like in practice. The Agent Request History for the Google Search Console connector shows the mix of context_store_search calls (pulling from the pre-synced store) and direct list calls:

Same for Google Ads, showing calls against ad_groups and campaigns:

Can I use Airbyte inside Claude?

Yes, you can. Run Airbyte Agent MCP inside Claude, ChatGPT, Cursor, whatever MCP client you want, and you get access to the Context Store and Airbyte’s library of 50+ agent connectors. There are more marketing connectors coming soon. I built this in the Airbyte UI, but you can build wherever you feel comfortable.

What does a run look like?

Here is the Sessions view after the latest run. One session, three connectors, 29 tool calls.

What are tool calls? They are direct calls to the APIs from the agent handled in Airbyte. That means no manual API wrangling, no pagination logic, no OAuth refresh on my end. Agent orchestrates the connector calls, runs the gap analysis, scores keywords, writes to Notion. Done.

Why Airbyte Agents specifically?

The connectors already exist. Google Search Console and Google Ads aren't native tools in Claude or any general-purpose LLM. Wiring them up directly means building custom MCP servers, handling OAuth, managing token refresh, dealing with pagination, maintaining it all when APIs change. Days of setup before your first report. On Airbyte Agents, both connectors are there out of the box. One auth step.

The actual key to this is the Context Store. The Context Store is a cached and searchable replica of a selected few entities from all your data sources that Airbyte manages and refreshes automatically for you. This means that the agent does not have to call the LLM with the raw response of the API, paginate and filter each one; rather, it just has to query a pre-filtered dataset.

Feel free to give it a try with 30days free trial.

I'm not an engineer. I am a marketer who just knows how to use AI. The Automations builder let me define this whole workflow without writing code. Described the job, attached connectors, set the schedule, live.

This pattern works beyond keyword gaps

The underlying workflow is: pull structured data from two systems, compare it, publish a report. You can apply that across marketing and ops, for all sorts of workflows.

I'm going to build more of these and share what I find. I’m excited to build new AI workflows for demand gen using Airbyte Agents.

Limitless data movement with free Alpha and Beta connectors

Introducing: our Free Connector Program ->