Put Your Brand in the Answer | GEO Product Notes 11: The Features I Cut

The features I cut: the hardest decisions in product evolution

New features are easy to see.

Another entry point, another dashboard, another automated capability: each gives the release notes something to say. Removing a button, delaying a capability, or telling users that the product cannot yet make a judgment automatically offers much less immediate satisfaction.

I also went through the stage where more features seemed to make the product more complete.

The question library, knowledge base, article generation, publishing, monitoring, opportunities, and retesting could all keep expanding. Whenever a new problem appeared, the natural response was to add another feature and cover it.

The deeper the GEO product moved into a real workflow, the more I found that some feasible features should not be implemented in the most convenient form. Some judgments a model can produce should not immediately become business conclusions. A button placed one step too early can make the user act before the evidence is strong enough.

The hardest question in product evolution changed from “What should we build next?” to “What should we delay, and what should we stop doing this way?”

Cutting a feature does not only mean deleting code. It can mean moving a capability out of the default path, adding a necessary threshold, or refusing to present an unreliable inference as a certain result.

01 First, I cut the idea that one score explains everything

Monitoring products naturally gravitate toward a summary score.

Mentions, rank, sentiment, citations, and source coverage can all become numbers. Weight them together and the page gets a prominent GEO score. The team can quickly compare this week with last week, or the brand with a competitor.

The score has not disappeared completely. It can still help sort and browse, and an impact score can indicate which opportunities deserve attention first.

What I cut was the idea of treating the summary score as the final conclusion.

The same score can come from completely different realities. The brand may be absent from natural questions but present in many brand-named prompts. Mentions may rise while critical product facts are wrong. Many citations may appear without any official source. A falling monitoring success rate can also make the result look worse.

If the score no longer lets a person drill down into questions, answers, sources, and collection environments, greater precision only produces a stronger illusion of certainty.

I now keep the score where it belongs: as a way to find something worth opening, not as a replacement for the team’s judgment.

That tradeoff taught me that a feature does not need to disappear from the interface to be cut. Reducing it from “answer” to “index” is a meaningful product contraction.

02 I cut complete article generation when there is no evidence

Content generation is easily optimized for completion rate.

Give the system a question and it should immediately return a structured, fluent article. When material is missing, a model is good at filling the gaps with general knowledge and industry language.

That kind of completion can hide the most important gap in GEO content.

An article may contain the brand name, a method, and a conclusion without a single confirmed brand fact that is approved for publication. It does not look like a failure. It may look more complete than a careful short draft, but it cannot answer why any of its brand claims are safe to state publicly.

I therefore added a hard threshold to article generation: the project must have at least one confirmed publishable fact. Searchable internal knowledge and a monitoring answer attached to an opportunity cannot replace that evidence.

Without it, the system no longer forces out a generic brand article. It asks the user to return to the project’s fact card and add a source and confirmation basis.

This lowers the percentage of clicks that immediately produce an article. In the real workflow, it removes a class of low-quality success that is especially hard to notice in time.

I would rather have the product stop clearly than give an unsupported sentence the appearance of public fact simply because it reads smoothly.

03 I moved “generate improvement article” behind the opportunity decision

After monitoring finds brand absence, a “Generate improvement article” button feels natural.

The user sees a problem and acts immediately. The path is short, and the loop looks complete. I was drawn to that smoothness in early versions of the opportunity page.

It relies on a hidden assumption: the system already knows that the opportunity is valid and that an article is the correct action.

In reality, one absence may be random variation. The question may be irrelevant to the brand. Existing content may already be sufficient while the monitoring environment has changed. Even when the opportunity is valid, the next action may be an official page, investigation of a wrong source, a revised question set, or more comparable samples.

In a later revision, I moved the action button behind opportunity confirmation.

The awaiting-decision stage first shows why the issue may matter, the original answer, whether the evidence persists, limitations, expected improvement, and success criteria. Article and publishing tasks become available only after an operator confirms that the opportunity should enter execution.

That deliberately adds friction to a frequently used feature.

The friction is not decorative approval. It forces the product and the person to answer one question before resources are committed: what are we actually trying to solve?

04 I cut automatic clarification based on a negative classification

Sentiment looks well suited to automation.

If an answer contains words such as risk, weakness, or not recommended, a model or rule can label it negative, create a high-priority opportunity, and recommend that the brand publish a clarification.

Words become misleading outside their context.

“Not suitable for large teams” may be a reasonable product boundary, not a brand crisis. A warning about an industry risk may not evaluate the target brand. A sentence with contrast, quotation, or comparison can be truncated by a simple rule into the opposite conclusion.

When an incorrect sentiment judgment directly triggers a public action, the product has moved from discovering a clue to speaking on the brand’s behalf.

In the current flow, an automatically detected negative is only a clue awaiting review. It becomes a formal opportunity only after a person returns to the full answer and confirms that it truly is a negative statement about the brand.

I did not remove automatic detection. I removed its power to become a conclusion by itself.

A model can narrow the review surface, but brand-reputation judgments cannot discard context and responsibility just because automation is faster.

05 Some intelligent-looking inferences remain deliberately unavailable

Another class of features was not removed because it was never allowed to reach the final step.

Suppose an answer cites several third-party pages and a competitor performs better on the same question. The system can calculate the difference between the two source sets. It still cannot reliably prove which competitor claim a page supported or that placing the brand on the same site would change the result.

Automatically creating an opportunity to “publish on this site” from a domain difference looks intelligent, but the causal chain is incomplete.

Factual accuracy has the same issue. The product can identify an answer that may discuss product capability, price, qualifications, or service scope. Deciding whether the statement is correct requires comparison with confirmed brand facts. Until the human review and correction chain is complete, I do not want the system to create correction opportunities at scale.

These are the hardest features to cut.

They can already display an appealing result and appear to be one step away from usefulness. When attribution is still missing, that final step is exactly the one the interface should not take on the user’s behalf.

Keeping the feature unavailable does not mean the direction lacks value. It means the product does not yet have enough evidence to accept the consequences of the action it would trigger.

06 I cut automatic improvement when action tasks finish

State machines invite a smooth story.

Create an article, finish review, publish it, and when every action task completes, automatically mark the opportunity improved. The team gets a satisfying progress line from discovery to resolution.

Publication only proves that an action happened.

It does not prove that AI accessed the page, began mentioning the brand, or changed its sources or factual language. Closing the opportunity when an internal task finishes substitutes internal progress for an external result.

I separated those states.

After an action finishes, the opportunity moves to awaiting verification. The system returns to the original question, platform, and collection basis, then compares valid post-action samples with the frozen baseline. Only meeting the success criteria defined beforehand can move it to improved. A missed criterion can make it ineffective or send it back to execution. Insufficient evidence remains unknown.

The loop is slower, but the word “improved” means something again.

07 I did not make publishing automation succeed at any cost

The deeper a browser extension goes, the easier it is to make full automation the goal.

Identify the page, fill the title, upload images, select a category, click Publish, and bring the link back. Once every platform has an adapter, the user seemingly no longer needs to enter the flow.

Running against real external platforms forced me to cut that idea.

An unidentified account, a changed page structure, a security challenge, a declaration that needs a decision, or a missing result after final submission cannot be solved by trying one more time. In particular, when the final step may already have created public content, an automatic retry can produce duplicates.

The extension must be allowed to stop.

If it cannot confirm the conditions before submission, it stops. If the final action may have happened without a response, it enters result needs verification. Steps involving account responsibility and platform declarations retain human confirmation. Automation can perform deterministic, recoverable actions. It cannot bypass uncertainty for a higher completion rate.

I did not cut publishing. I cut the belief that full automation is always more mature than human takeover.

08 I now ask four questions before deciding what to cut

These tradeoffs led me to a review method that does not feel like a conventional feature review.

First, does this feature help the user see a fact, or does it manufacture a conclusion for them?

Second, if the judgment is wrong, can the result be recovered easily? A ranking hint can be recalculated. An article published to the wrong account cannot simply be made to have never happened.

Third, does a smooth experience hide insufficient evidence? The more a feature promises one-click completion, automatic attribution, or a universal score, the more closely I examine what it removed from the middle.

Fourth, does it move the user closer to a real next step? Another page, entry point, or chart does not necessarily reduce decision cost. It can create another information silo in the workflow.

If a feature amplifies false certainty, triggers an external action that is hard to reverse, or makes the same information require maintenance in several places, I now prefer to narrow it first.

Product restraint is not building fewer features. It is placing automation where the evidence and responsibility can support it.

09 After the cuts, the product looks more like a way of working

These decisions share one thing: each gives up some immediate sense of completeness.

Without evidence, the article does not generate. Before an opportunity is confirmed, the action button does not appear. Before a negative signal is reviewed, the system does not automatically demand a clarification. When tasks finish, the page does not immediately announce an improvement.

This kind of product does not always give the fastest or most polished result.

It does separate observation, judgment, action, and verification more clearly, and it preserves “we do not know” when that is the truth. That boundary matters more in GEO than another automated button because the system works with changing AI answers, external platforms, and brand facts rather than fixed inputs inside a closed system.

I used to think of product evolution mostly as accumulating capabilities. I increasingly see that it also means knowing which capabilities should not be handed directly to a model, which steps need human judgment, and which results must wait for outside evidence.

Those tradeoffs lead to a more specific question: when AI gets a brand wrong, what should the product trust and preserve, and who should make the correction?

The next note explains how evidence, audit history, and human review form a correction path that does not rewrite history.