Bradley here.

I've seen thousands of local business websites where the written content is carefully optimized and the images are basically treated like decoration.

IMG_4827.jpg.

IMG_4828.jpg.

DSC_1049.jpg.

Alt text: "image."

Fantastic.

Meanwhile, the client has a folder full of actual photos showing their technicians, projects, equipment, offices, vehicles, and real work in real locations.

That's useful data.

Computer vision gives agencies a way to process that information at scale. But there's a right way to automate image optimization, and there's a very stupid way.

Let's get into it.

IN THIS EDITION

QUICK HITS

THE DEEP DIVE

Here's the problem.

A local business might have 500 photos sitting in Google Drive.

Nobody knows what's in half of them.

The filenames are garbage. The alt text is missing. The images aren't organized by service, location, project, or date.

And the agency is creating another 1,000-word blog post about why the client is "the trusted local expert."

That's backwards.

The client already has evidence.

You just aren't using it.

Computer vision models can inspect an image and identify visual information.

For a local HVAC company, that might mean recognizing:

Now imagine doing that across 5,000 images.

A human team could spend days cataloging those photos.

A vision model can create an initial classification almost immediately.

That's where automation becomes useful.

This is where I've watched agencies screw things up.

They take the output and turn every image into something like:

That's garbage.

Alt text exists to describe the image.

If the image shows a technician replacing an air conditioning capacitor, say that.

If the image shows a completed commercial rooftop unit installation, describe that.

If the image is purely decorative and provides no useful information, don't force a giant SEO paragraph into the alt attribute.

The objective is accurate context.

Not keyword stuffing.

If I were setting this up for an agency today, I'd use a five-step process.

Step 1: Build the image inventory.

Pull together the client's website images, GBP photos, project photos, staff photos, and approved media assets.

Give every image a unique identifier.

Step 2: Run computer vision classification.

Have the model identify what is actually visible.

Objects.

People.

Equipment.

Environment.

Actions.

Potential service relationships.

Don't publish anything yet.

This is your first-pass dataset.

Step 3: Add business context.

This is where the SEO team earns its money.

The model might see:

The business data might tell you:

Those are two very different levels of useful information.

Combine visual data with verified business information.

Now you can build a much stronger description.

Step 4: Generate the supporting metadata.

For each image, you can create appropriate:

Again, don't publish blindly.

Have a human review the important assets.

Step 5: Connect the image to the right page.

This is the part agencies miss.

An image shouldn't exist in isolation.

If you've got a real photo of a commercial HVAC installation, put it on the relevant commercial HVAC page or project case study.

If you've got a real photo from a specific service area, use it where that geographic context is actually relevant.

The image becomes another piece of evidence supporting the page.

This is where I think image optimization gets much more interesting in the AI-search era.

Search engines are trying to understand entities and relationships.

Real imagery can contribute to that broader picture.

I've talked before about the importance of real-world evidence and localized image data. The important point is that the image should support a genuine business relationship, not manufacture one.

A real photograph of a client's technician performing a real service is fundamentally different from a stock image selected because it contains the keyword "plumber."

One is evidence.

The other is decoration.

Here's another opportunity.

A local business doesn't just need to tell Google where it operates.

It needs to consistently demonstrate that relationship across its digital footprint.

If a client has legitimate photos from jobs in their service area, those images can become part of that geographic story.

But don't start inventing locations.

Don't add a city to an image because you want the page to rank there.

Don't claim a project happened somewhere it didn't.

That's how you turn a useful system into a pile of bullshit.

Use verified information.

Real photos.

Real locations.

Real services.

That's the semantic approach.

The real advantage here isn't that AI can write an alt tag.

That's a five-second task.

The advantage is being able to process thousands of assets consistently without assigning an employee to manually inspect every image.

Think about a multi-location client.

You've got:

You can build an image inventory, classify the assets, identify gaps, and prioritize the images that actually deserve optimization.

Now you've got a repeatable fulfillment process.

That's where AI becomes useful to an agency.

One more thing.

Optimization for optimization's sake is bullshit.

If an image doesn't help the page, don't add it just to increase the image count.

If the image is blurry, irrelevant, duplicated, or useless, delete it.

If the image is valuable, make it accessible, descriptive, properly named, appropriately compressed, and connected to the right content.

The goal is a better information system.

Not a bigger media library.

Your clients are sitting on years of first-party visual evidence.

Most agencies aren't doing anything with it.

That's the opportunity.

Instead of generating another generic image with AI, ask the client for the photos they already have.

Then use computer vision to help organize that mess.

That's a much more defensible asset.

PRODUCT & SERVICE SPOTLIGHT

This is where Geo Niche Edits fit into the broader strategy.

When a client needs both local and vertical relevance, I want those two signals working together.

The Service Catalog describes Geo Niche Edits specifically for situations where a business needs combined local and vertical relevance. The service is designed to layer Entity-Based Relevance and Geographic Signaling together.

That's important here because image optimization is only one layer.

Then you still need the external web to reinforce what that business is, what it does, and where it operates.

That's where relevant geo-niche signals fit.

»» Order Geo Niche Edits ««

AWESOME APPS

Most "AI stacks" still ship every prompt to someone else's API. That gets expensive, and it means client content leaves the building. FreeToken is an open-source engine for running huge Mixture-of-Experts models locally on hardware you already own.

Here's why it's worth watching:

The Agency Lever: If you're building internal research agents, content pipelines, or client-data workflows, local inference cuts API burn and keeps drafts off third-party servers. This is infrastructure, not a Canva button. Treat it that way.

I personally use FreeToken with Qwen3.8-Flash-Next NVFP4 as my daily driver and run about 60% of all agency operations through it. It’s secure and saves loads in token costs.

»» Check out FreeToken here ««

WINS OF THE WEEK

AI is making it possible for small agencies to build fulfillment systems that would have required custom development a few years ago.

Image classification is a perfect example.

The technology isn't the hard part anymore.

The hard part is designing a workflow that combines machine speed with human verification.

I teach this and other agency automation in the Semantic Mastery MasterMIND.

THE HIDDEN WIN

Here's what most agencies are overlooking:

Your client's old photos might be more valuable than the next piece of AI-generated content.

Go ask for the photo archive.

You might find hundreds of images documenting actual work that has never been properly connected to the website.

That's first-party evidence.

Organize it.

Describe it accurately.

Put it on the pages where it actually belongs.

Then make sure the rest of the web reinforces the same business, service, and geographic relationships.

That's a hell of a lot more defensible than publishing another stock-photo blog post.

One of the strongest principles in our existing work is that real-world evidence matters.

The same philosophy applies to images.

A real photo from a client's job site can demonstrate experience in a way that another paragraph of generic marketing copy can't.

And when that image is properly organized and connected to the relevant service and location, you've created another useful data point for the broader entity.

That's the play.

Stop treating images like decoration.

They're data.

Use computer vision to process them at scale.

Use humans to verify them.

Then connect the visual evidence to the entities, services, and locations that actually matter.

Build better content.

Build better relevance.

Bradley Benner
Founder, Semantic Links