From Keywords to Data | Evolution of Search at ISC Education

In 2026 my pages ranked higher than ever, readers stayed longer than ever, and the traffic still fell. Meanwhile the work I am proudest of never shows up in a traffic chart at all.

Vu Quynh Anh - SEO writer, then SEO lead, then AI-search and data · ISC Education, 6/2021 - now

Two lines that should not go together

A note before you read. I have signed an NDA with ISC Education. So all of ISC’s data in this article is hidden. You will see which way things moved, but not the numbers behind them.

In September 2026 I pulled the summer numbers for isc.education and compared them with the same summer a year before.

The average position had gone up by several places. The best it has ever been.

The clicks had gone down, sharply.

Five years ago those two lines could not happen together. If you ranked higher, you got more traffic. That was the deal. This year (and I think the years after as well), the deal broke, and this article is about what I did when it did.

Five years, three titles

• 6/2021 - 11/2023. SEO Content Writer, freelance.
• 11/2023 - 10/2025. Senior Search Engine Optimizer.
• 11/2025 - now. AI-Search Optimizer. In practice, mostly data.

ISC Education is a Vietnamese study-abroad consultancy. A family walks in with four questions. Which country. Which university. Which course. Which job after graduation.

The last one is the one parents care about most, and the hardest to answer. So it needs a clear meaning. In the Australia dataset I built, a graduate job is not a line from a brochure. It is a chain of four checkable facts. The career the university says the course leads to. The official Australian occupation that career matches. The states that nominate that occupation for skilled migration, and the visas it opens. And what it typically pays.

For example: Aboriginal and Torres Strait Islander Health Worker, nominated by Victoria and Western Australia, with a typical annual salary attached.

Scholarships are a separate thing. They come after the four questions are answered. They are a reward on top of a decision, not the reason for it. Nobody picks Australia over the UK for a discount.

For four years my job was to answer those questions on the website. I wrote hundreds of articles doing it. For the last year my job was to answer them in a dataset, so a consultant could answer them in front of the family.

The first job is now judged by a number AI is eating. The second job has no number at all. Both are the best work I have done.

Did the content get worse?

That is the first question anyone asks when traffic drops. So I checked it three ways, in two tools. Same summer, 2025 against 2026:

• Average position: up. How Google rates the page.
• Impressions: about the same. How many people asked.
• Clicks: down. How many came.
• Engagement rate: slightly up. Did they stay.
• Time on page: up. How long they stayed.

Sources: Google Search Console for the first three. GA4, Organic Search, for the last two. Visitors decline cookies and are not counted in either year.

Position went up. Position is Google’s own score for the page. A worse page drops. Ours climbed.

Impressions held. That is the demand. Students are asking the same questions, as often as before.

The readers who arrive stay longer. A worse page loses people faster. Ours keeps them longer. And organic search is still the most engaged big channel on the site, ahead of paid search, social and direct.

Take the brand name out and nothing changes. Better position, same impressions, fewer clicks.

So the content did not get worse. By every measure the content controls, it got better. The only thing that fell is the one step it does not control. The click.

Where the clicks went

Here is one question, asked in dozens of different ways. Which university is number one in the world?

A year ago we ranked near the top for it, and it was one of our best sources of clicks. This year the impressions are almost the same. The position slipped a little. And the clicks nearly disappeared.

What happened is that the answer is one word, and Google now writes that word at the top of the results page. The student reads it and leaves. My article is still there, still ranked, still correct. Nobody needs to open it.

Neither tool has a column called “lost to AI”. Search Console and Analytics are not fully built for AI reporting yet. So I cannot give you a final number. What I can give you is a page that ranks better, is seen as often, holds its readers longer, and is visited less. On that results page, only one thing changed. It was not my article.

What AI-search optimisation actually is

My title changed to AI-Search Optimizer in November 2025. People ask me what that means. It is not SEO with a new name.

A search engine sends a person to your page. An AI assistant reads your page and answers the person itself. So the goal changes. You are no longer trying to be clicked. You are trying to be the source the AI chooses to quote.

That changed my work in two ways.

I researched questions, not keywords. Students do not type into ChatGPT the way they type into Google. They write long, personal questions. My GPA is this, my family can pay about that much. Is the UK realistic? That is a question you would ask a consultant, not a search box. So I collected those questions and wrote briefs from them.

I made each page easy to quote. Put the fact at the top, in one plain sentence. Name the source. Put a date on every number. Make it clear which university, which intake, which currency. An AI can only quote a fact it can find and trust.

Did it work? After the first round, Gemini started citing the domain within days. And this year Analytics shows a new channel called AI Assistant. Most of it comes from ChatGPT, then Gemini, with a little from Perplexity and nothing yet from Copilot.

The numbers are small. But look at the quality. These visitors are more engaged than visitors from Google search or from paid ads. Gemini visitors stay the longest of all. When an AI sends someone, it sends someone who wants the detail.

Then I looked at where they land. This surprised me.

ChatGPT sends people mostly to the homepage, the events page, the contact page, and articles like study-abroad consultancies in Ho Chi Minh City. Gemini sends them to the best study-abroad consultancy, consultancies for Australia, for the US, and to a UK visa guide.

So the AI is not quoting us for single facts, like a fee or a ranking. It is quoting us when a student asks “who should help me?” and “how do I do this?”. It seems to recommend ISC as a company, and ISC’s guides as a process.

That is what AI-search optimisation comes down to for a consultancy. Two jobs:

• Make the company easy to recognise. Who ISC is, where the offices are, what it does, which countries. Stated the same way everywhere, so the AI can describe it with confidence.
• Make the process pages the best on the web. Visa steps, application steps, how to choose a course. Long answers, with sources and dates. Those are the questions AI still sends people to read.

And the tools are not ready either. Search Console’s report on generative AI features is still in beta. The AI Assistant channel in Analytics is new. Neither tool is fully equipped for AI reporting yet. So everything about AI in this article is provisional, not final. Read it as an early signal.

Why “less credit” is the real problem

A report reads clicks. A manager reads the report.

For four years my work and my chart moved together. This year they split. Rankings up, demand flat, engagement up, and a traffic line pointing down. To explain that, you need a meeting and five rows of data.

That is what “less credit” means. Not that nobody noticed the work. That the one number everyone looks at no longer measures it.

And if AI is going to read our answers instead of our readers, then the answer underneath matters more than ever. Which is where the second half of this year went.

From keyword list to schema

In my last year at ISC, most of my time was not SEO. It was data.

It came straight out of the SEO. After four years I knew every question a Vietnamese student types about studying abroad. How much does it cost. Is my IELTS enough. How long is it. When can I start. What can I do after.

The consultants needed answers to exactly those questions. But the answers were spread across hundreds of university websites, in hundreds of formats.

So I took the keyword list and turned it into a schema. Each question became a field.

• How much? → tuition fee, in one currency
• Is my English enough? → IELTS score
• How long? → duration, in years
• When can I start? → intake dates
• What level, what subject? → study level and subject area, mapped to one standard list
• What can I do after? → careers, official occupations, nominating states, visas
• Any money off? → a scholarships table, joined to each university

An SEO keyword plan and a consultant’s dataset are the same list of questions. One is answered with an article. The other is answered with a row.

The pipeline

Nobody at ISC had built one before. I built it as a standard ETL pipeline, in four stages.

• Acquisition. Crawl every course and scholarship page and store a raw copy on disk before parsing anything. The raw layer never changes, so every later step can be re-run on the same input.
• Extraction. Parse the raw copies into structured records. Course, level, fee, IELTS, duration, intake, source URL.
• Transformation. Normalise every field to one format. Standardise labels to one list. Match scholarships and campuses to the right university. Deduplicate courses listed on several pages. For Australia, enrich every programme with careers, occupations, states, visas and city living costs from government sources.
• Loading. Export to two workbooks a consultant can filter by subject, budget and English score.

Every record keeps its lineage: the URL it came from, and for any derived field, how it was derived.

What came out was two datasets. The UK one covers well over a hundred universities and tens of thousands of courses. The Australia one covers the country’s universities, their campuses, their programmes, and the jobs and visas those programmes lead to.

But size is not the honest part. The honest part is how full each field is. I will come back to that.

I thought the hard part would be acquiring the data

Before I started, I was sure acquisition would be the fight. Blocked crawlers. Pages that only render in a browser. PDFs. Sites that change while you crawl them.

Some of that happened. But it was the part I could plan for. A crawler fails, you fix it, you run it again. By the end, acquisition was just time.

But, the actual hardest part was: To check it

Testing was a nightmare. I do not have a softer word for it.

Checking sounds like one job. It is two.

Did we get every row? That is record completeness. If a university offers a certain number of courses, the dataset needs exactly that many. Not fewer because one faculty used a different page layout. Not more because one course appeared under three campuses. Missing rows are the dangerous ones, because nothing looks wrong. A course that is not there does not throw an error.

Did we get every metric out of each row? That is field completeness. A row can exist and still be half empty. The fee is on the page, but in a table the parser skipped. The IELTS score is there, but inside a sentence: “6.5 overall with no band below 6.0”. And every empty cell asks the same question. Did the university not publish it, or did I not extract it? In a spreadsheet those look identical. They mean opposite things.

Then the part that made it a nightmare. It was the first year. There was no previous version to diff against. No baseline. Nobody who knew the right count. When the pipeline gave me a number, the only way to know it was right was to open the website and count.

So I did, by hand. Pick a university. Count its courses on the site. Compare. Open ten rows. Check every field against the page. Find a fee in the wrong column. Fix the parser. Re-run on the raw layer. Check again. Next university.

In year two you diff against last year and only look at what moved. In year one, everything is new, so everything needs a look.

Unexpected: university sites are big messes

I assumed universities would be the tidy part. They are the messiest sources I have ever worked with.

They write a printed prospectus, then paste it onto the web. No schema. No structured data. No API. Every university invents its own way to say the same thing.

Australian programmes used dozens of labels for what is really five or six levels of study. UK scholarship values came in hundreds of different wordings.

And often the answer is simply not there. A large share of courses have no fee, the first number every parent asks. Some universities do not publish one per course. Some hide it in a PDF. Some show it only after you choose a campus and a start date. Australian scholarship deadlines are almost never published in a form you can extract.

It is the same thing I saw from the SEO side, from the other end. The information students need is out there. It is written by people who aren’t equipped to maximize search and school awareness on the Internet.

What I left behind

Some things I did not finish. I want to say them plainly.

• Fields are still empty, only up to around 100% filled.
• Some jobs are borrowed. When a university did not list careers for a course, I took them from a similar course at another university. They are marked, but still need more research
• Nobody checked my checking. It was the first year. There was no old version to compare with. So every number was checked once, by hand, with nothing to check it against.
• I could not prove the AI loss with one number. I can show the pattern. Better rank, same demand, fewer clicks. Google does not give me the exact number.
• The data has a new owner. I handed everything to my boss. Next year’s version will have something to compare against. I will not be the one who uses it.

What I’d tell myself a year ago

On SEO

Do not show the traffic chart alone. Put rank and time on page next to it. On its own, a falling line looks like failure, even when the work got better.

Stop chasing questions with a one-word answer. “Which university is number one” lost almost all its clicks in a year. “What is a PhD” gained. Google now answers the short ones itself. Write for the questions that need a whole page.

Start the data work sooner. The most useful thing I made this year was not an article. It was a row.

On data

Write the columns before you open a single website. I already had them. They were my keyword list.

Know the right count before you collect anything. Write down how many courses a university has before you run anything. Otherwise you cannot tell if you got them all.

Give every empty cell a reason. “Not published” and “not found by me” look the same in a spreadsheet. They are not the same.

Expect checking to take longer than building. In the first year there is nothing to compare with. So everything gets checked by hand.

Mark every guess. A guessed row with a mark is honest. A guessed row without one is a mistake waiting for a parent to find it.

If you are starting in SEO now

The market is saturated. Anyone can publish an article, and AI publishes them faster. And as this whole article shows, even a good article now gets fewer clicks every year.

So here is where I would go. Data.

You do not have to leave SEO to get there. You extend it. You already know what the market asks, and that is the hardest part of building a dataset. Most data people do not have it. Add crawling, parsing, normalising and validating on top, and you become the person who turns what people search for into something a business can run on.

That is the next level of the job. In my experience it is the next level of the salary too.

What’s new in Search Console

If you stay in SEO, relearn Search Console. It changed a lot over the last year, and most people still only open the Performance tab.

Here is what is there now:

• Insights. A simpler view of how your content is doing. What is growing, what is trending, which pages people find. Useful for a quick weekly check without building a report.
• Recommendations. Short cards on the Overview page that flag problems for you. For example: a page recently got fewer impressions than usual, with the size of the drop. It tells you where to look before you notice it yourself.
• Other properties. You can now watch more than your website. Your YouTube channel and social media profiles can sit next to the site, so you see your whole search presence in one place.
• Generative AI features (beta). A report on how your site performs inside Google’s AI features. It is the first sign of Google measuring the AI side at all. It is still in beta, which is exactly why the AI findings in this article are not final.

None of this replaces your own analysis. But a year ago, none of it existed. Search Console is slowly admitting that a click is no longer the only thing that matters. Learn the new sections now, while they are still small.

In conclusion

Five years in, my best work does not only show up in the traffic chart. It shows up in a dataset a consultant can filter in seconds. I would rather be judged on that.