Articles on AI agents, automation and custom software

Most small shops start on Shopify, and that's usually the right call. You pick a theme, add your products and you're selling by the weekend. Nobody has to think about servers.

But Shopify was designed around boxes: stock, shipping, sizes, returns. If what you sell is a download, like an eBook, a course or a set of templates, you're living in a house built for someone else. For a while that's fine. Then, little by little, it isn't.

This article is for owners who have started to feel that. It covers when Shopify is still the right answer, the signs that it has stopped fitting, and what moving off it really takes. Along the way we'll use one of our own projects as an example: Sessosophia, an Italian shop that sells eBooks, which we moved off Shopify this year.

01When Shopify is still the right answer

Don't move if:

• You sell physical things. Stock, shipping and returns are what Shopify does best, and rebuilding them is expensive.
• You change your products all the time and want to do it yourself, without asking anyone.
• There's nobody to look after a custom shop once it's built.

If that's you, keep your shop where it is and spend the money on better photos and better words. It's what we'd tell you on a call, too.

02Three signs it's time to look again

1. Your shop looks like everyone else's

A theme decides how every page looks before you've written a single word. You can change the colours and the font, but underneath, your shop has the same bones as thousands of others.

For some products that doesn't matter. For others it matters a lot. When people buy something personal, such as advice, a course or a book about their own life, they want to feel there's a real person behind it. A grid of products can't give them that.

Sessosophia sells guides on a very personal subject. On a standard theme, it looked like it could have been selling anything.

2. The admin is "manageable"

Nobody leaves Shopify because the admin is broken. They leave because it's manageable: everything is possible, and nothing is quick.

An admin built for every kind of shop is full of things a digital shop never touches, like shipping zones, stock levels and parcel weights. Meanwhile the questions you actually ask are spread across different screens and apps. "Did she get her download?" can mean checking the order, the customer, the email app and the file app before you know.

A custom admin starts from those questions instead. Sessosophia's team described the old one as manageable, not easy. The new one answers their everyday questions on one screen: what happened to an order, a button to resend a lost email, refunds, and who hasn't downloaded yet.

3. You're paying for a lot you never use

A platform plan is priced for everything the platform can do. A small digital shop uses a sliver of it, pays for all of it, and then pays again for add-ons to fill the gaps.

When the shop is your own code, the bill covers what it actually uses: hosting, email and card payments. For Sessosophia, running costs fell by 70 to 90%.

03"Can't we keep Shopify behind a new design?"

You can. It's often called going headless: a custom front end, with Shopify still running behind it.

It fixes the first sign. It doesn't touch the other two. You keep the same admin and the same monthly bill, plus a new site to look after. If the admin and the cost are part of why you're unhappy, a new front end only gets you halfway.

04Where do you stand?

One "worth a look" on its own is normal. If both apply and the add-ons line does too, the platform has probably stopped fitting.

05What moving actually takes

Moving off Shopify doesn't mean building a shop from nothing. Payments still run through a proven provider (Stripe, in Sessosophia's case). What gets built is the part that's yours: the site and the admin.

1. Start with what you sell. Write down your products, how they're grouped and how people buy them. For a digital shop that list is short, and everything Shopify carries for boxes simply never gets built.
2. Bring your history with you. Old orders come across, so nothing disappears on launch day. Sessosophia's Shopify orders sit right next to the new ones.
3. Design the admin around real questions. Ask the people who will use it what they look up every week, then give each answer a home.
4. Bring the emails with you. Order confirmations, download links and cart reminders were Shopify's job. On your own shop they need to be set up again, and you can add new ones, such as a reminder for buyers who haven't downloaded yet.
5. Leave a guide behind. Whoever runs the shop should be able to do their job without phoning a developer. We wrote Sessosophia's in two languages.

06What you give up

To be fair, a custom shop comes with trade-offs:

• Someone has to look after it: updates, backups and the odd bug. Budget for that, whether it's your own team or a partner.
• Bigger changes go through a developer. Sessosophia's catalogue rarely changes, so this works well for them. If yours changes every week, think twice.

07So, should you move?

If you ship boxes, probably not. Shopify was built for you.

If you sell something digital, ask yourself three things. Does your shop feel like yours? Does "manageable" quietly eat part of your week? And how much of your monthly bill pays for things you never touch?

If the answers make you wince, it might be time. And if you'd like a second opinion, we're happy to look at your setup and tell you honestly whether moving would be worth it.

Picture a buyer with a real problem and a budget. Ten years ago they typed a few words into Google, opened four tabs and started comparing.

Today many of them open ChatGPT, Gemini, Perplexity or Copilot and ask a full question: "What's the best AI automation agency for a midsize UK firm?"

They get back a confident paragraph that names three companies, says a line about each, and sometimes cites a source or two.

If your brand isn't one of those three names, you weren't outranked. You were left out of the conversation entirely.

01Three eras of being found

SEO (search engine optimisation) was about the ten blue links. Rank higher, get the click. The page did the persuading.

AEO (answer engine optimisation) arrived with featured snippets, People Also Ask and voice assistants. The goal shifted from ranking to being the answer: one source quoted directly at the top of the page.

GEO (generative engine optimisation) is where we are now. AI engines don't quote one page. They read many, combine them, and write their own answer. Sometimes they name brands and sometimes they don't. They cite sources, but not always yours.

These eras don't replace each other, they stack. Many AI answers are built from search results: Copilot draws on Bing, and Google's AI Overviews draw on Google. That makes good SEO the foundation, and often a better ranking is the first sign of a better AI result. But ranking alone no longer guarantees you get mentioned.

02The problem: everyone's guessing

Most "AI visibility" advice today rests on a screenshot. Someone asks ChatGPT a question once, sees their brand (or doesn't), and draws a conclusion.

That approach has three problems:

1. AI answers vary. Ask the same question five times and you can get five different shortlists.
2. Engines change constantly. Models and search back ends update without notice, so this week's answer isn't comparable to last month's.
3. No baseline means no proof. If you can't say where you started, you can't prove that anything you did made a difference.

At Pilotree we kept hitting the same wall. Clients asked a simple question, "Are we showing up in AI, and is it getting better?", and the honest answer was "we can't tell yet." So we built a way to tell.

03How we measure it: Squawk

Squawk is our internal search and AI visibility platform. It measures how visible a brand is in Google and in AI answers, plans and logs the work that improves both, and proves the result with numbers a client can check for themselves.

The method follows a few rules.

1. Ask like a buyer, and ask more than once

Every client gets a set of 20 to 40 questions, written the way a real buyer would phrase them and grouped into topics. They fall into four kinds:

• Brand: "What does [brand] do?" Does AI describe you correctly?
• Category: "Best [service] for [buyer] in [place]". Are you named when the buyer doesn't know you yet? These are the questions that win new business.
• Comparison: "[Brand] vs [competitor]". How are you positioned against rivals?
• Control: a real question on a topic we will not work on. This is the yardstick (more on it below).

Each question is asked five times on every engine: a ChatGPT stand in, Google AI Overviews, Gemini, Perplexity, Copilot and Claude. Every answer is stored in full, with its date, model version and citations, so every number can be traced back to a stored answer.

2. One headline number

The Squawk Score (0 to 100) is the weighted share of AI answers that mention you, counted on category and comparison questions only. Brand questions always name you, so they're left out. Control questions are the yardstick, so they're left out too.

A mention only counts if it's really you. If an AI talks about a different company with a similar name, that's logged as a confusion, not a mention. Confusions are a useful finding in their own right.

3. Separate real change from noise

This is the part most AI visibility reports skip.

Each metric is reported with a noise band, a 95% range. A change smaller than that range isn't claimed as a win.

The control questions then catch the engines changing under you. Say your category questions rise 12 points and the control questions, which nobody touched, rise 5. The engines moved by roughly 5 points on their own, so the lift you can actually claim is 7.

When the control moves beyond its own noise band, the report says so plainly and holds off on claiming a win.

4. Find out who AI cites instead of you

For every topic, Squawk lists the domains the engines cite and sorts them into six types: own, competitor, editorial, community, reference and directory. The mix decides what to do next.

• Competitor pages are winning? Build or improve your own pages on that topic.
• Editorial or community sources are winning? Earn coverage there. Another page on your own site won't help.

04What actually moves the needle

The good news is that most fixes improve SEO and GEO at the same time:

1. Deeper pages on the topics buyers actually ask about. Thin service pages lose to competitors who answer the full question.
2. A named, credible author or expert on the content.
3. Structured data that makes it easy for machines to read who you are, what you offer and where.
4. Consistent facts across the web. If your site, directories and press coverage disagree about what you do, AI repeats the confusion.
5. Presence where AI looks: industry press, community threads and the directories that keep showing up in citations.

None of this is new. What's new is being able to show that it worked.

05Intelligent growth means measured growth

The future of digital marketing isn't SEO or AEO or GEO. It's all three, measured together, on the same topics, with a fixed method and a clear before and after.

Rankings still matter. Being the answer still matters. But the question that decides the next decade of discovery is simpler:

When your buyer asks AI, does it say your name?

If you don't know the answer, that's the place to start.

Squawk helps companies get found in Google and in AI answers, and proves it with numbers they can check. Want to know whether AI names your brand? Get a free visibility check.

Your next customer might never see your homepage. They ask ChatGPT for "the best payroll tool for a 40-person startup", read a paragraph that names three companies, and click one. If you're not in that paragraph, you weren't in the running.

That shift has produced a pile of acronyms. SEO, GEO and AEO get sold as three separate disciplines, often by three separate agencies. They aren't separate. They're one question asked in three places: when someone looks for what you do, are you the answer?

We built Squawk to measure this for clients. What follows comes from those audits.

01What the three words mean

In practice people use AEO and GEO loosely, and they overlap. Don't lose sleep over the labels. The useful part is the last column: three ways of being found, all of which now matter.

02Four myths worth ignoring

"SEO is dead." It's the opposite. Many AI answers are built on top of search results. Google's AI Overviews are written from Google's own index, and Copilot answers from Bing. When an assistant searches the web before it answers, the pages it finds are mostly the pages that already rank. A better ranking is often the first sign of a better AI result.

"GEO needs a whole new kind of content." Mostly it doesn't. The pages AI engines like to quote are the pages people like to read: they answer the question in the first paragraph, use real numbers, cite sources, and say who wrote them. Most of the fixes that move AI answers also move rankings.

"Add an llms.txt file and you're done." llms.txt is a proposed file that lists your key pages for language models. It's cheap to add and harmless, but no major engine has confirmed that it uses it. Treat it as tidying, not strategy.

"You can track your ranking in ChatGPT like a Google position." Ask ChatGPT the same question five times and you can get five different answers, naming different companies in a different order. A single screenshot proves nothing in either direction. More on measuring this properly below.

03The checklist: what to do this week

These are the checks that most often fail when we audit a site, roughly in the order we'd fix them.

1. Let the AI crawlers in

Each engine has its own crawler, and some have two: one that builds an index and one that fetches pages live when a user asks. The ones worth checking in your robots.txt:

• OpenAI: GPTBot, OAI-SearchBot, ChatGPT-User
• Anthropic: ClaudeBot, Claude-User
• Perplexity: PerplexityBot, Perplexity-User
• Google: Google-Extended, which governs Gemini's use of your content. It doesn't affect Google Search, and AI Overviews read your site through the normal Googlebot.

Then check that your site actually serves them. robots.txt can say yes while your CDN or firewall says no. Bot-protection rules increasingly block AI crawlers by default, and Cloudflare started doing that for new sites in 2025. Fetch a key page with each crawler's user agent and look at the status code. A 403 means that engine can't read you, whatever your robots.txt says.

2. Put the content in the HTML

Many crawlers don't run JavaScript, or run it badly. If your page's text only appears after a script loads, some engines see an empty page. View the page source (not the inspector) and check that your main content is there. Content hidden in tabs and accordions is also easy to miss, so keep the important bits visible.

3. Write answer-first pages

For each question a buyer might ask, there should be one page that answers it well. The patterns that get pages quoted:

• Answer first. A short opening paragraph, under about 60 words, that directly answers the question in the title.
• Facts and figures. Prices, timelines, numbers and named examples rather than adjectives.
• Sources. Link the studies, standards or official bodies behind your claims.
• Tables for anything comparable: plans, specs, options.
• A named author and a visible "last updated" date.
• An FAQ with matching FAQPage structured data, where it genuinely fits.

4. Make the web agree about who you are

When an AI engine describes your company, it leans on what the wider web says, not just your site. Three things help it get you right:

• A Wikidata entry for the company, with your official website on it.
• sameAs links in your site's Organization structured data, pointing at your real profiles: LinkedIn, Crunchbase, your business listing. Use the same set on every page.
• Facts that match everywhere. Your name, founding year, location, services and contact details should be identical on your site, LinkedIn, your business listing and the main directories. Every mismatch is a chance for an engine to pick the wrong one.

A related problem is namesakes. If another company shares your name, AI answers can blend the two. Consistent entity data is your best defence.

5. See who gets cited instead of you

Ask the questions your buyers ask, and look at which sites the answers cite. Sort them roughly into competitors, editorial sites (press, industry publications), community sites (Reddit, forums), reference sites (Wikipedia, docs) and directories. This tells you what kind of work will help:

• Competitors' pages win: build or improve your own page on that topic.
• Editorial or community sources win: another page on your site won't help. You need to be mentioned there: coverage, guest pieces, honest answers in the threads people actually read.

04How to tell whether any of it worked

This is where most "AI visibility" reporting falls apart. Answers vary from run to run, and the engines change their models whenever they like. Say a number goes from 20% to 28%: that might mean your work paid off, or it might mean nothing at all. A defensible method has five parts.

1. Use the questions buyers actually ask. Write 20 to 40 of them in your buyers' words, and split them by kind:

• Brand ("What does Acme do?") checks whether AI describes you correctly.
• Category ("Best payroll tool for a 40-person startup") is the one that wins new business.
• Comparison ("Acme vs Rival") shows how you're positioned.
• Control questions are on a topic you deliberately won't work on. They're your yardstick.

2. Ask each question several times on each engine. We use five samples per question per engine for a baseline, and run the baseline twice, about a week apart, before any work goes live. The two baselines show you how much the numbers move on their own.

3. Report every number with its noise band. Give each metric a 95% range, worked out by resampling the questions it's built from. For a single question the range is a Wilson score interval: 3 of 5 answers is 60%, anywhere from 23% to 88%. A change only counts if it's bigger than the range. If it isn't, say so plainly.

4. Subtract the control. If your category questions rise 12 points and the control questions, which you never touched, rise 5, the engines shifted under everyone. Your real improvement is 12 − 5 = 7 points. If the control moves beyond its own noise band, flag the period as "engines changed" and don't claim a win from it.

5. Freeze the method. Lock the questions, competitors, engines and model versions before you collect data. Compare only like with like. Change the method and you need a new baseline.

None of it is advanced statistics. It's what a before-and-after claim needs to hold up when someone checks it.

05One score, because it's one problem

Once you measure all three together, the overlap becomes obvious. One crawl of your site serves the search audit and the AI audit. A deeper page, a named author or a fixed crawler rule moves rankings and AI answers at the same time. That's why Squawk folds them into a single score: how often you're named in AI answers, read alongside your search ranking and how readable your site is to AI, each with its noise band.

You don't need a platform to start, though. Run the checklist above, ask your buyers' questions five times each, write down what you see, and do it again in a month. You'll learn more from that than from any dashboard of single screenshots.

If you'd rather have it measured properly from day one, Squawk does exactly this, and the first check is free.