Here’s how GA4 tables in BigQuery are structured, three first SQL queries — by day, by page and by source — and how to avoid overpaying for data processing.

Here’s how tag gateway works in plain words, what it really gives you, what it doesn’t do, how to turn it on via Cloudflare and how to check the result.

Here’s where to check them, how to read PageSpeed Insights and Search Console results, and why the numbers differ between tools.

Here’s how to set up all three records for Google Workspace email on your own domain: what to add, in what order and how to check the result.

Here’s how the reCAPTCHA types differ and which one to choose for a form, where to create a key now, where to find the secret key for a plugin and what happens if you exceed the limit.

Here’s how to get the map code, how to make it work well on phones, how to embed directions — and when you can’t do without a Google Maps Platform key after all.

Here’s how to check conversion tracking yourself: where to see the statuses, how to make a test conversion, how to compare the numbers with real enquiries and where data usually gets lost.

Here’s how to build your first report from scratch: connect data, add four useful blocks, set up the date range and share it with colleagues.

Here’s how to check consent mode yourself in 10–15 minutes — three ways, from the simplest to the most precise — and the mistakes that come up most often.

Here’s the installation step by step: create an account and container, add the code, set up your first tag, test it in Preview mode and publish.

Here’s the installation step by step: create a property and data stream, add the tag to your site in one of three ways, and check that data is actually coming in.

Here are three tools that answer different questions, how to read their results and the errors that come up most often.

Here’s where to find the reason, which issues come up most often, how to fix them and how to speed up the re-review.

Here are the verification methods, how to prepare for one of the most common — a video recording — how long the review takes and what to do if Google says no.

Here is the whole process step by step: which property type to choose, how to verify ownership, what to do straight after verification and the small things people most often get stuck on. With access to your domain at hand, it takes 15–20 minutes.

Here is who these visitors are, where they trip up most often and what is worth fixing on your site right now — without a rebuild and without chasing trendy protocols.

This is not a roundup of “the best website speed test tools”: one server, one Windows machine and honest impressions after the first measurements.

Why people come to us

• The site takes ages to open and visitors leave. On a phone the page needs several seconds to load. Some people close it before it appears, and the ad that brought them there has already been paid for.
• A PageSpeed report is a wall of red numbers. LCP, CLS, TBT: it is unclear what they mean, how serious they are and what to fix first.
• Your developer says “it is fast for us”. On their powerful computer it is. On a customer’s phone with a mobile connection, it is not.
• The site got slower after an update. New photos, a banner or a plugin made the page heavier, and nobody noticed for months.
• It is hard to tell what to pay for. Someone offers “website speed optimization” for serious money, and you have no way to check what will change and by how much.
• An online shop and Google rankings. Slow product pages and a slow checkout push customers away before they pay, and Google takes page speed and comfort into account when ranking sites, especially on phones.

What an assistant connected to PageSpeed Insights can do

An ordinary AI chat cannot open your website or measure it. It gives generic advice such as “compress your images”. A connected assistant runs a real Google test and works with the results. You can ask in your own words:

• “How fast does my site open on a phone?” — it tests the page and answers plainly: fast, acceptable or slow, and why.
• “What slows the site down the most?” — it finds heavy images and unnecessary scripts and styles that keep the page from appearing right away, and sorts the fixes by importance.
• “Which of our pages is the slowest?” — it tests up to ten pages in one go and ranks them from worst to best.
• “Did the fixes help?” — it compares measurements before and after, so you can see whether your contractor delivered.
• “How do we compare with a competitor?” — it compares two pages by speed: yours and theirs.
• “What do real visitors experience?” — if the site gets enough traffic, it shows data from real Chrome users; if not, it says honestly that there is no data.
• “Is everything else fine?” — the same report also scores accessibility for people with low vision, basic SEO and adherence to Google’s technical best practices, and the assistant explains those scores too.

How it worked on our own site

Before launching the service, we tested our own website. On a phone the home page scored 66–71 out of 100 and took about five seconds to load fully, while the SEO and best-practice scores were 100. Within a couple of minutes the assistant named the cause: ten layout and script files that the browser has to download before it can show the page, together delaying it by almost two seconds. It also found unused code (about 100 KB of styles and 50 KB of scripts) and showed that our SEO page opens noticeably faster than the home page. The lesson: pages of the same site can differ a lot, so it is worth checking more than the home page. That kind of review used to take an hour in reports. This time a few minutes were enough.

What you get

• Answers instead of reports — a conclusion, the numbers and a clear next step.
• A fix list sorted by importance that you can hand to your own developer or to us, with the expected gain for every item.
• Control over contractors — a measurement before and after the work shows whether the site really got faster.
• Early warning — you can repeat the check after every site update and catch a slowdown at once, not six months later.
• Savings — Google’s service is free, so you do not pay per audit. Only the assistant subscription costs money, and it costs less than a single speed audit from a specialist.
• No access to your site needed — only public pages are tested, exactly as any visitor sees them. We do not ask for passwords, admin access or your analytics.
• Your data stays with you — the program runs on your computer, and the Google key is registered in your name.

What you need for the installation

You do not need a powerful computer or a separate server. The AI runs in the cloud, and your computer only runs a small connector program between the assistant and Google’s service.

• A computer with Windows 10 or 11, macOS or Linux.
• 8 GB of RAM and 2 GB of free disk space.
• A stable internet connection.
• A paid subscription to the Claude AI assistant in your name. The basic plan is enough.
• A free Google key for the speed service. We create it together in about ten minutes.
• Your website address, and nothing else.

How the setup works

Three formats

• Remote. We connect to your computer through a secure remote-access session, and every action is visible on your screen. For companies across Poland and the EU.
• On site. We set everything up at your place, on the computer where the assistant will work. Białystok and the Podlaskie region; other cities by arrangement.
• In your cloud. If your company works in a cloud environment, we install it there, so several employees can use the assistant.

Steps

1. A 20-minute call: which pages matter most to you and what worries you about speed.
2. The Google key: a free technical key that works only with the speed service. Your Google password is never shared with anyone, and you can switch the key off at any moment.
3. Installation and connection, usually about an hour.
4. A first measurement together on the same day: the home page and 5–10 key pages, and you leave with a fix list.
5. Short training and a list of ready-made questions for every week; optionally, a repeat measurement after your fixes.

The honest limits

• The assistant changes nothing on the site by itself: it measures and advises. The fixes are made by a developer. If you want, we can take them on as a separate job.
• Google’s measurements are lab tests on a typical mid-range phone, so a result can vary by a few points from run to run. That is why we look at the trend and the specific causes, not at a single number.
• Not every site has real-user data: with little traffic, Google does not collect it.
• Only public pages are checked. Anything behind a password, such as a customer account or the checkout after login, is invisible to the assistant.
• AI can make mistakes in its explanations, so next to every conclusion it shows the measurement numbers, which you can double-check on the official PageSpeed Insights site.

Frequently asked questions

Do you need access to my website?

No. Only public pages are tested, exactly as any visitor or search robot sees them. Passwords, the admin panel and analytics are not needed.

How is this better than opening PageSpeed Insights myself?

You get the same service, and it is free. The difference is that the assistant reads the English report with its dozens of metrics for you: it explains it in plain language, remembers earlier measurements, compares them and sorts the fixes by importance.

Will it work if my site runs on WordPress, Shopify, Wix or Joomla?

Yes. The finished page is tested the way a browser receives it, so the platform does not matter.

Will my site get faster by itself?

No. The setup measures and advises, and the speed-up comes from the fixes. What you gain is knowing which fixes will make a real difference and which are minor, so you do not overpay.

How much does it cost per month?

Google’s service is free. The only recurring cost is the AI assistant subscription, about 20 US dollars a month for the basic Claude plan. Our setup is a one-time fee.

How it works under the hood

The assistant connects to PageSpeed Insights through an MCP server, an open standard that gives AI assistants access to external services. Before installing it, we check the module for safety, as in our safe rollout of AI agents and MCP servers service. How we set this up on our own site is covered in the blog post “How I connected PageSpeed Insights to Claude with MCP”.

Tell us which pages of your website take too long to open — we will show you what the assistant can tell you from your own measurements.

• Answers instead of charts. You ask in plain words — "how many enquiries came in last week", "where do our clients come from", "why did traffic drop after the site was rebuilt" — and you get an answer you can act on, not a table of figures to decode.
• A regular report without your involvement. Weekly or monthly, whichever we agree on, a summary arrives in plain language: what changed and what deserves your attention.
• Time saved. No opening Analytics yourself, hunting for the right report and working out what its terms mean — the assistant does that for you.

How it works

The AI assistant gets read-only access to your website statistics in Google Analytics — viewing only, with no right to change anything on the site or in the analytics itself. From there you simply talk to it: ask it to compare months, work out the conversion rate from one traffic source, or explain why a particular page lost visitors.

What it saves you

An in-house marketer or analyst who goes through site statistics regularly costs thousands of euros a month. A one-off agency report once a quarter is out of date within a week. The AI assistant answers a specific question immediately rather than on a schedule: the setup cost is fixed, and after that you simply use what is already running, with no monthly invoice for every question.

What you need for the setup

• An ordinary computer with internet access — Windows or Mac, no special hardware requirements.
• Your Google account with access to the site's Google Analytics (or a willingness to grant read-only access to our working account).
• A subscription to an AI assistant such as Claude or ChatGPT — if you do not have one yet, we will help you choose and set it up.

No separate server or costly infrastructure is needed: everything runs on your computer or in the cloud you already use.

How the rollout works

Most clients prefer it remotely: we arrange a call, set up the access and the assistant over a remote desktop session — usually one working session, without needing your constant involvement. If you would rather do it in person, we come to you in Białystok and the surrounding area and configure everything directly on your computer.

Technically the connection is made through MCP (Model Context Protocol) — the standard way to give an AI assistant safe access to exactly one service, in this case read-only Google Analytics statistics and nothing else of yours. If you also want the assistant to see search data, this pairs with our service for connecting Google Search Console to an AI assistant. Tell us what you want to know about your website — we will scope the work and the timeline before we start.

• Reads your inbox and sorts it by meaning. Adverts, newsletters and client enquiries stop sitting in one pile — the agent sees the difference and labels messages accordingly.
• Finds the mail that is easiest to lose. An enquiry from your website form, a reply to a job advert, a question from a client — if it landed in spam or sank among newsletters, the agent pulls it back into view.
• Drafts the reply, not just a notification. For a routine enquiry the agent prepares a draft answer — you check it and send, instead of writing from scratch.
• Never decides for you quietly. Deleting, sending, moving an important message — only with your confirmation. The agent shows what it found and what it suggests doing.

What you get out of it

• You stop losing clients. An enquiry that used to sit unread among spam for a week reaches you the same day.
• You save time every day. No more scrolling through dozens of messages to find the two or three that matter.
• You reply faster than your competitors. A client who gets an answer within the hour, rather than in three days, is less likely to go elsewhere.
• You can see what happens to your mail. A clear report of what the agent found and what it did — no black box.

What it needs to run on

No separate server is required. The agent runs on an ordinary office computer or laptop — the one your employee already works at, or a dedicated machine if that suits you better. All it needs is a stable internet connection and access to your mailbox. If your mail is already in the cloud — Google Workspace, Microsoft 365 or your hosting provider's mail — the agent works through it directly, so there is no extra infrastructure to buy.

How the setup works

• Remotely. We configure everything over a secure remote session — faster and usually cheaper. Available to companies across the EU.
• On site. We come to your office, set it up on the spot and show you how it works right away. Białystok and the Podlaskie region, other cities by arrangement.

In both cases the only thing with access to your correspondence is the agent configured personally for you — not a third-party service and not a shared account. Your mail stays on your machine and in your mailbox, which matters for GDPR: we do not move company correspondence into any intermediary service.

Technically this connection is made through AI agents and the MCP protocol — a secure translator between the AI model and your mail software. If the security side is what concerns you most, we offer a separate service: reviewing and safely rolling out AI tools.

There is also a September plot twist: Google reformed the way API access is granted, so half the guides online were already stale on the day I installed this — and I walked the obsolete path in full before I noticed.

What people come to us with

• Ads burn the budget and bring few enquiries. The campaign targets "business services" while the customer types "how much does a company website cost" — the money goes on clicks that lead nowhere.
• The website speaks industry language, not customer language. The services page is written the way professionals talk, so Google never shows it for the questions people actually ask.
• A new service launches with demand unchecked. First you pay for the website and the campaign, then you find out whether anyone was looking for it.
• It is unclear which words are worth paying for. Some phrases cost pennies, others sit in an expensive auction, and which of them actually brings a customer is invisible.
• Budget leaks on clicks that never convert. Searches with "free", "DIY" or "jobs" bring people who will never buy, and nobody has built a negative keyword list.
• It is unclear which language to speak to the market in. In Poland the same service is searched for in Polish, Ukrainian and English — and the split is rarely what you would guess.
• The agency sends a keyword list with no explanation. Where the numbers came from, and whether they mean anything, is impossible to verify on your own.

What the assistant can do once it is connected

A plain AI chat cannot see search statistics and picks phrases from general reasoning — elegant and beside the point. A connected assistant takes the numbers from Google's own tool and answers from them. You ask in ordinary words:

• "What do people search for when they want a website built in Bialystok?" — collects the real phrasings and groups them by topic.
• "How many searches a month does this phrase get, and is it seasonal?" — shows a year of search volume and flags the months when demand rises.
• "Which words are expensive and which are cheap?" — reports competition level and the approximate cost per click for each.
• "Which language do people use to search for my service in this city?" — compares demand by language and region so you are not translating the site on a hunch.
• "What should go into this service page?" — suggests the phrases people really use and flags the ones not worth fighting for.
• "Which negative keywords should I add?" — isolates the searches that bring non-buyers.
• "How many clicks can I expect on this budget?" — builds a monthly forecast of impressions, clicks and spend.
• "What do people search for around a competitor's page?" — takes their URL and derives related queries.

What the result looks like

You do not get a raw export from an ad panel. You get a working document: a list of phrases, each with monthly search volume, competition level and an approximate cost per click, and next to it a short verdict on where that phrase belongs. Some go into headings and page copy, some into ad campaigns, and for others the advice is to leave them alone — competitors overpay for them and they bring few customers.

Decisions from a document like this look unglamorous, which is exactly why they save money. A phrase with a dozen searches a month and an expensive click is not a priority, however well it reads in your service description. A phrase hundreds of people search for, with no decent page from any competitor, is a reason to write that page first.

How it worked for us

We installed this setup for ourselves first. It earns its place not when you want to admire charts, but at the moment of a decision: write this page or not. The assistant checks the phrasings, discards the invented ones and shows what people genuinely ask. That is how the service pages on this site were written — including the one you are reading.

We also walked the whole path through Google's access procedures, and it took longer than the installation itself: in September 2026 Google changed how API access is granted, and half of the guides online stopped being accurate. The full account is in "Wiring Google Keyword Planner into AI". For you it means we take the administrative part on ourselves.

What you get

• A list of phrases with numbers — your customers' searches, monthly volume, competition, approximate cost per click.
• A clear "what goes where" — which phrases belong in page copy, which in ads, and which to leave to competitors.
• A sanity check before you spend — you see whether anyone searches for a new service before paying for a site and a campaign.
• Negative keywords — so the budget stops leaking on "free", "DIY" and job hunters.
• Oversight of contractors — any agency keyword list can be checked against the same Google data.
• Savings — the planner itself is free; the only recurring cost is the assistant subscription, less than an hour of a PPC specialist per month.
• Your data stays with you — everything runs on your computer and in your Google account, and access can be revoked at any moment.

Who benefits most

• Service businesses in Poland and the EU — you see which language and which city your service is actually searched in.
• Online shops — which product searches grow before the season and how people name the product.
• Local businesses — demand in your city and district rather than "the national average".
• Anyone launching a new line — demand checked before money goes into a site and ads.
• Anyone already paying for ads — a review of the words your budget goes to.

What you need for the installation

No powerful machine, server or graphics card is required: the AI runs in the cloud, and what sits on your computer is a small go-between program connecting the assistant to Google.

• A computer running Windows 10 or 11, macOS or Linux.
• 8 GB of RAM and 2 GB of free disk space; the program itself takes about 30 MB.
• A stable internet connection.
• A Google Ads account — free, and you do not need to run campaigns to research keywords; if you have none, we will create it together.
• A paid Claude subscription in your name — the basic plan is enough.
• Application access to Google's data — the paperwork and verification are on us; all you do is confirm in Google's own dialogue.

How the setup runs

Three formats

• Remotely. We connect to your computer through a secure remote session — every action visible on your screen. For companies across Poland and the EU.
• On site. We set everything up in your office, on the machine the assistant will run on. Bialystok and the Podlaskie region; other cities by arrangement.
• In your cloud. If your company works in a cloud desktop environment, we install it there, so several employees can use the assistant.

Steps

1. A 20-minute call: which business lines, cities and languages need checking.
2. Google-side access is handled by us: we create the account used for keyword research if there is none, and pass the application verification. Your Google password is never shared — you grant consent in Google's official dialogue and withdraw it whenever you like.
3. Installation and connection on your computer — usually an hour or two.
4. The first analysis on your own topics — together, the same day.
5. A short walkthrough and a list of ready-made questions to ask before every new page and campaign.

Honest about the limits

• The assistant never launches or changes anything in your ad account: it reads planner data and advises. Campaigns, bids and budgets stay under your control.
• Google estimates demand approximately — it is guidance for decisions, not a headcount of real people or a guarantee of enquiries.
• If the Google Ads account has never spent anything, Google reports volumes as ranges — "100 to 1K per month". For choosing a direction that is usually enough.
• For very narrow topics Google has little data, so the conclusions are more cautious — better to see that now than after a paid campaign.
• Access to the data is granted by Google after it verifies the application: it is not an instant switch, and we build that time into the schedule.
• AI can be wrong, so every conclusion comes with the numbers behind it, which you can re-check in Google Ads yourself.

Frequently asked questions

How is this different from keyword research done by an agency?

An agency delivers a spreadsheet once, and six months later it is out of date. Here the tool stays with you: demand can be checked before every new page, service and campaign without commissioning the work again. The numbers are the same — they come from Google's planner.

Do I need to be running Google Ads already?

No. A free Google Ads account is enough and no campaign has to be live. Keyword research is most useful before the advertising starts, so the budget is not spent blindly.

Can I look at demand for a specific city and language?

Yes. The planner separates data by country, region and city, and by the language of the query. For a business in Poland that often decides everything: demand in Polish, Ukrainian and English can differ several times over, and differently in each niche.

Is it safe to give AI access to an ad account?

The assistant has exactly three planner actions available: generate ideas, show search history and calculate a forecast. This module cannot create campaigns, change bids or spend budget — the capability is simply not in it. You grant consent in Google's dialogue and revoke it in your account settings.

Will it work if the site runs on WordPress, Shopify or Joomla?

Yes. The data comes from Google rather than from the site, so the platform is irrelevant. There may be no site at all yet — demand is checked before you order one.

I already have an SEO contractor. Will this get in the way?

The opposite: you will be able to check their proposals against the same Google data and talk specifics — why this page is needed and why that phrase is not worth the budget.

What does it cost per month?

Google's planner is free. The only recurring payment is the AI assistant subscription — around 20 dollars a month for the basic Claude plan. Our setup is a one-off, from €90.

How soon is the first result?

The first analysis of your topics happens on installation day. A detailed phrase list with recommendations is usually ready within two or three working days.

How it works under the bonnet

The assistant reaches the planner through an MCP server — an open standard that lets AI assistants use external services. We use a compact open-source module that can only read planner data, and we review it for security before installing, the same way we do in "Safe rollout of AI agents and MCP servers". How we set it up for ourselves, and what went sideways, is described in "Wiring Google Keyword Planner into AI".

This service pairs well with its neighbours: Search Console with AI shows which queries already find you, Google Analytics with AI shows what visitors do next, and the speed check shows whether slow pages are losing them.

Tell us which business and which market to analyse — you will have an estimate before any work starts.

This is not a “best MCP servers” roundup. One server, one Windows workstation, a site in four languages, and honest impressions after the first days of use.

We build client sites on different platforms — WordPress, Joomla and others — so I have nothing to defend. What follows is an honest comparison: where WordPress wins so clearly there is nothing to argue about, where Joomla is genuinely stronger, and why the choice of platform affects the outcome far less than people assume.

Below is an honest account of that setup: what it consists of, why one channel turned out not to be enough, what it genuinely speeds up and what it does not, and the two incidents it cost us along the way. No talk of revolution or transformation — only what is running on a production site right now.

Below is a checklist without the legal phrasing: what exactly belongs on the site, how the requirements differ for a sole trader and for a limited company, why an "Accept" button with no "Reject" button counts as a breach, and how all of it looks on a live example — this very site.

Below is a way to check this yourself in half an hour, with no tools and no agency: five questions to ask about your own market, what to write down in the answers, and what to do with the result. Plus an honest look at why you are not being named, and how much of that can actually be fixed.

Below is the technical side: who is covered and from when, why February 2026 affects even the smallest firms, how the 10,000 zł threshold is really counted, where the NIP field belongs in a checkout form, and how the B2B path differs from B2C.

Below is the practical breakdown: who the law covers and who it does not, and why the exemption does not always hold; what it actually demands of a website once you strip away the language of standards; the five things sites most often fail on; and how to check your own site in fifteen minutes without hiring anyone.

Below is a breakdown of a real case: what the Japanese keyword hack looks like from the inside, why restoring a backup does not help, which detail almost convinced me the cleanup had failed, and how to check your own site in five minutes. The owner agreed to let me tell this story with the site named — for which I thank him separately: stories like this are usually hidden, and that is exactly why other site owners never hear about them.