Key takeaways
- General models cover the 30 percent of performance marketing that is platform knowledge. They cannot access the 70 percent that is your account history and decisions.
- Prompt engineering helps for a single session but does not solve the problem. Context typed in goes stale, does not accumulate and cannot be prioritized by business impact.
- ChatGPT, Claude and Gemini each have genuine strengths for ad work. None of them maintain your account context between sessions by default.
- Almond's Business Context Engine captures tests, results and decisions continuously, keeps them current and ranks them by relevance so recommendations reflect this week's reality.
- Almond supports bring-your-own LLM, so you can run the same general models on top of a context layer that knows your account rather than choosing between them.
General AI models are genuinely impressive for performance marketing. ChatGPT, Claude and Gemini can explain bidding strategies, write ad copy variants, audit campaign structures and reason through attribution models. That capability is real and worth using.
The problem is not what they know. It is what they cannot know, and why that gap matters more than most marketers realize.
The 70/30 split that defines performance marketing
Great performance marketing is roughly 70 percent business context and 30 percent ad platform knowledge. The 30 percent is well documented: bidding logic, audience segmentation, creative best practices, campaign structures, match types, conversion tracking setup. General models have absorbed most of this from public documentation, industry blogs and platform guides.
The 70 percent is different. It is the reasoning behind why the account is structured the way it is. The test you ran in March and killed because creative failed, not targeting. The KPI your team dropped at the last quarterly business review. The budget shift that happened on Monday. The competitor that launched last week and changed the entire conversation in your leadership meeting.
None of that is written down anywhere a general model can reach. It lives in Slack threads, meeting recordings, email chains and three people's heads. A general model has no access to it, and no mechanism to acquire it.
What a side-by-side looks like in practice
Consider a concrete scenario. Your Meta campaigns are showing a rising cost per acquisition. You open a general model and ask what is driving it.
A capable general model will give you a solid answer. It might walk through audience saturation, creative fatigue, iOS signal loss, auction competition, landing page mismatch or seasonal demand shifts. The reasoning is sound. The framework is correct. If you are new to Meta, it is genuinely useful.
Now consider what the answer is missing. It does not know that you paused your best-performing creative set two weeks ago because Legal flagged a claim. It does not know that you shifted budget toward a new product line that has a structurally higher CAC because leadership decided to prioritize it. It does not know that you already tested broad audience expansion in Q1 and it underperformed by roughly 40 percent against your segmented approach. It does not know that your attribution window changed last month.
So the model recommends broad audience expansion. You already killed that test. It suggests revisiting the creative you paused. Legal still has not cleared it. It flags the CAC increase as a problem to solve when your team already knows it is an accepted trade-off for the current quarter.
The answer is not wrong by general standards. It is wrong for your account, your team and this week.
Why this is structural, not a prompt engineering problem
The instinct is to fix this by writing better prompts. Type more background. Paste in the brief. Summarize the last QBR before asking the question.
This works, partially, for a single session. It does not scale for several reasons.
First, the context you type in is only as current as the last time someone typed it. Budget shifted Monday. The brief you pasted is from March. The model optimizes against stale information and has no way to know it is stale.
Second, context typed into a prompt window does not accumulate. Every session starts from zero. The test you ran and killed in March is not in the next session unless someone types it again. Learning does not compound.
Third, volume and relevance become a problem as you add more sources. Paste in everything and the model weighs a throwaway Slack recommendation the same as the CEO's strategy call. More context without prioritization produces noisier answers, not better ones.
The underlying issue is that general models are stateless by design. They are built to answer questions from whatever context fits in a prompt window. They are not built to hold an account's history, stay current as decisions happen and rank what matters this week against what is just background noise.
What ChatGPT, Claude and Gemini each position themselves as
It is worth being precise about how these tools describe themselves, because the comparison is more nuanced than a simple ranking.
ChatGPT, based on its public materials, positions itself as a general-purpose AI assistant with strong reasoning and code capabilities. OpenAI has added memory features and custom GPT configurations that can hold some persistent context, though these are user-managed and session-scoped. Verify current capabilities at openai.com.
Claude, based on Anthropic's public positioning, emphasizes long context windows and careful reasoning. Its 200,000-token context window is genuinely useful for processing long documents in a single session. It does not maintain context between sessions by default or connect to live data sources unless integrated via API. Verify current capabilities at anthropic.com.
Gemini, based on Google's public materials, positions itself with strong integration into Google Workspace and, in some versions, access to real-time search. For Google Ads specifically, its knowledge of the platform is deep given Google's role in producing that documentation. It does not have access to your account's decision history or cross-platform context by default. Verify current capabilities at deepmind.google.
All three are excellent at the 30 percent. None of them, in their standard configurations, can address the 70 percent.
The compounding cost of the gap
The practical cost shows up in a few consistent patterns.
Repeated tests. A team asks a general model for optimization ideas and gets back a suggestion they already ran and abandoned. Nobody typed that history into the prompt, so the model does not know. The team either runs the test again or spends time explaining why it will not work, which defeats the purpose of asking.
Stale strategy. The model's recommendations are calibrated against whatever context was last loaded. If leadership changed the primary KPI three weeks ago and nobody updated the brief, the model is still optimizing for the old one. This is not a model failure. It is a context maintenance failure that gets attributed to AI being unreliable.
Senior review tax. Because a general model cannot be trusted to know whether its output is true for this account right now, a senior person has to check every recommendation against account reality. The efficiency gain from faster generation is partially offset by the review cost. For agencies in particular, this is where the margin improvement that AI was supposed to deliver has not materialized.
Key-person concentration. The only person who can prompt a general model usefully on a given account is the one who already holds its context in their head. The tool that was supposed to reduce dependency on key people has deepened it, because the context lives in the person, not the system.
How Almond approaches the same problem
Almond is built as a Business Context Engine, not a replacement for general models. The distinction matters.
Almond reads your ad accounts, Slack, Notion, Sheets, email, meeting notes and CRM in read-only mode, connected in roughly four minutes. It captures every test, result and decision as it happens, keeps that context current as things change and compounds it over time. Context is classified by relevance, recency and business impact, so a strategy decision from the CEO's Monday call ranks above a throwaway Slack thread from last Tuesday.
The result is that when you ask a question, the business context is already there. You do not type the background. The model does not recommend the test you killed in March, because Almond knows you killed it and why.
On the model question specifically: Almond supports bring-your-own LLM. If your security team has approved Claude or GPT-4o or Gemini, you can run Almond's context layer on top of those same models. The general model handles the 30 percent it is genuinely good at. Almond supplies the 70 percent it cannot know on its own. The two are complementary, not competing.
Security is built in from the start. Context is never used to train models, is client-scoped so one client's data cannot surface in another's recommendations, is versioned and auditable, and is encrypted in transit and at rest.
Where Almond fits
If you are a performance marketer, media buyer, growth lead or agency founder who already uses general AI models and finds the output useful but not quite trustworthy for your specific account, the gap is almost certainly the context layer, not the model.
General models are not going to get worse at platform knowledge. That 30 percent will keep improving. The 70 percent, your history, your decisions, your tests, your business context, is not something a general model can solve by getting smarter. It requires a system designed to capture, maintain and prioritize that context continuously.
Almond's early access is open. Every member gets a detailed one-on-one Meta and Google ad account audit with the CEO, recorded. Waitlist is at almond.ai and requires a work email.
Frequently asked questions
- Can ChatGPT manage Google or Meta ad campaigns?
- ChatGPT can reason about campaign structure, bidding strategy and creative direction based on general platform knowledge. It cannot access your live account data, know your past test results or stay current with decisions your team made this week. It is useful for frameworks and drafts, less reliable for account-specific recommendations without significant manual context input.
- Is Claude good for performance marketing?
- Claude's long context window makes it useful for processing lengthy documents like creative briefs or audit reports in a single session. Based on Anthropic's public positioning, it does not maintain context between sessions or connect to your ad accounts by default. It handles general platform knowledge well but lacks access to your account's decision history and current business context.
- How is Almond different from using a general AI model for ads?
- Almond is a Business Context Engine that captures your account history, past tests, decisions and business data continuously and keeps it current. General models start each session from zero and cannot know what your team decided on Monday. Almond also supports bring-your-own LLM, so you can run the same models you already use on top of a context layer that knows your account.
- What is the biggest limitation of using Gemini for Google Ads?
- Gemini has strong knowledge of Google Ads based on Google's own documentation. The limitation is not platform knowledge but account-specific context. It does not know why your account is structured the way it is, what you tested and killed last quarter, or how your strategy shifted after last week's leadership call. That context has to be supplied manually every session.
- Can I use Almond with ChatGPT or Claude instead of replacing them?
- Yes. Almond supports bring-your-own LLM, meaning you can run Almond's context layer on models your security team has already approved, including GPT-4o and Claude. The general model handles platform reasoning and language generation. Almond supplies the business context those models cannot access on their own. The two layers are designed to work together.
Almond 2.0 early access
Get a 1:1 Meta and Google audit with the CEO.
Every waitlist member gets a 1:1 audit of their ad accounts. Work email only.