From setup to your first fixed error in one day.
Three steps for you. Underneath: an expert engine, a learning layer trained on results, and our own language model. Here is exactly how it works - no jargon, but no shortcuts either.
Three steps
Connect your accounts safely
Authorize via official Meta and Google OAuth. Phanes reads the performance data. Your passwords are never accessed or stored.
⏱ about 3 minutesReview your audit and tracking diagnosis
Get a health score, issue locations, and financial loss amounts. Fix tracking first, eliminate wasted spend second, scale performance last.
⏱ a few minutesApprove fixes or enable automated care
Every edit requires your approval. In the PRO plan, hand over routine tasks to Phanes within strict limits you define.
⏱ daily, in the backgroundPhanes technology
A dedicated expert engine, not a generic chatbot.
Standard AI agents interpret raw figures on the fly and make assumptions. Phanes runs on coded expert rules, full account context, and closed-loop verification.
Phanes
- Leverages full account historyFactor in seasonality and historical benchmarks from your specific account, not web averages.
- Coded expert knowledgeOver 200 proven optimization rules with strict thresholds. No improvised advice.
- Pinpoints root causes, not symptomsInstead of "improve CTR", you get the exact broken element, why it failed, and the supporting data.
- Differentiates seasonality from errorsA seasonal drop is a valid reason to hold steady, preventing unnecessary panic edits.
- Tracks its own performanceEvery applied change is evaluated against real sales performance a few days later.
- Safe change executionDry-run testing, API validation, diff previews, and required approval before any execution.
Generic AI agent / chat plugin
- Sees only a fractionLimited to whatever fits inside a single chat context window.
- Relies on generic adviceDispenses generic tips without specific rules or performance thresholds.
- Reports surface symptomsGenerates surface-level notes like "test creatives" without backing evidence.
- Overreacts to routine dipsUnaware of your business seasonality, it treats normal market dips as critical bugs.
- Ignores the actual resultsHas no mechanism to check if its suggestions improved performance or made it worse.
- Guesses numbers and changesCalculates metrics on the fly and edits settings directly without validation or error logs.
Six layers, one answer
We are not an AI company. We are an engine that learns from ad results.
Artificial intelligence has one job here: speaking plainly. Everything that is calculated, decided and verified is done by a deterministic engine - and every conclusion is settled against the account's real results.
Data
The collector pulls the full account history through the official Meta, Google Ads, GA4, GTM and Search Console APIs and stores it in daily snapshots: campaigns, ad sets, ads, search terms, placements, hours, setting changes. Every later decision traces back to a specific row of data.
Knowledge
Hundreds of rules from paid-ads expert practice and official platform documentation, written in a form the engine reads directly: condition, threshold, allowed action, rationale. Thresholds where sources disagree go into a catalogue of theses and are settled by data, not opinion.
Rule engine
Deterministic: the same data gives the same answer, every time. Rules run in order from measurement (can the numbers be trusted at all), through account structure, down to individual ad sets and search terms. The output is a list of findings with the loss priced in money and the rule cited.
Learning layer
Every proposal becomes a hypothesis with a predicted effect and a settlement date. A few weeks later the engine compares the prediction with reality and raises or lowers the rule's confidence. A cause miner searches the account history and checks which settings actually preceded a drop or rise in cost per result - and it, not intuition, sets the priorities for the next day.
Execution
A change in the account always follows the same path: proposal → simulation without writing → diff preview → your decision → write via API → journal entry. New campaigns are created as paused drafts, budgets change in percentage steps, deletion is blocked.
Security
Data encrypted at rest (AES-256-GCM) and isolated between clients. Access tokens never reach logs or responses. One client's data is never visible to another.
Learning from results
Phanes gets better every day.
Every day brings newly settled hypotheses: what really lowered the cost of a sale, a lead and a click on your account - and which settings drove it. Expert knowledge is the starting point; your results refine it. When there is too little data, the engine says so instead of guessing.
Our own language model
AI has one job here. And it runs with us, not with someone else.
The Phanes language model runs on our own compute infrastructure and is tuned to a single domain: performance advertising. We do not depend on the pricing or decisions of any external AI provider - your data never leaves Phanes, never trains someone else's models and is never passed on. That is why the cost of running an account is flat and predictable, and the subscription does not grow with your spend.
What the model does
- Translates the engine's conclusionsTurns a list of findings and numbers into sentences a business owner understands: what happened, what it costs, what to do.
- Understands instructionsIn chat it turns your sentence ("raise the brand campaign budget by 15%") into a concrete engine action - with the same preview and approval as any other change.
- Writes draftsAd copy, headline variants, reports and audits - based on your account data and in your language.
- Explains causesWhen cost per result rises, it describes what changed in the account and on the market, drawing on the cause miner's findings. A new version of the model, tuned on current expert knowledge and settled hypotheses, ships to Phanes every quarter.
What the model never does
- It doesn't calculateEvery number comes ready from the engine. It doesn't estimate, round or "recall" results.
- It doesn't decideWhether a proposal is safe and justified is ruled by the engine's policies. The model cannot bypass them.
- It doesn't write to the accountIt has no access to the Meta or Google APIs. The only write path goes through the execution track with your approval.
Rules that cannot be switched off
Account safety is built into the engine, not into the terms of service.
Changes only with your approval
By default you see every modification in a preview and approve it yourself. Full Care is something you switch on deliberately, within a spend cap you set.
Budget in steps
Budget changes are measured in percentages: a warning above 20%, a block above 100% in a single step.
New entities as drafts
New campaigns and ad sets are created paused. You launch them with one click, whenever you want.
Nothing gets lost
Permanent deletion is blocked - Phanes pauses, never deletes. Every change has a journal entry with date, author and diff.
Try Phanes on your own campaigns.
7 days of the full PRO version for free. No card and no commitment.
Try free for 7 daysBack to the home page