How we use AI for SEO (and what it cannot do)
AI is leverage, not a writer.

Most advice about using AI for SEO assumes the AI writes the articles. That is the version that does not work. On upgrid.ch, a brand-new Swiss domain with no backlinks and no budget for buying any, we used AI on almost every part of the project and it never wrote a page that anyone published. It built the pages, it pulled the tariff figures from official sources, and it compressed the research that normally eats the whole retainer. What that bought was time for editorial judgement, and the judgement is what earned the rankings.

The numbers below come from the Google Search Console API for upgrid.ch, not from an estimate tool. The full breakdown of the project lives in the Upgrid SEO case study.

What AI for SEO actually did on this project

It built pages, it did not write them

Upgrid sells into a technical Swiss category: electricity pricing, self-consumption communities, cantonal rules that change per municipality. The content that wins there is not prose, it is comparison tables, calculators and municipality-level breakdowns. Those are engineering jobs. AI turned a two-week interactive build into an afternoon, which is the difference between one useful tool per quarter and one per sprint.

It grounded every figure in an official source

Tariff numbers came straight from ElCom's official data, pulled through its API into the page rather than typed in from a model answer. That single decision removes the biggest failure mode of AI content: confident numbers that nobody can trace. If a figure appears on an Upgrid page, it exists in a federal dataset, and it updates when the dataset does.

It compressed the research, not the thinking

Reading Swiss energy regulation, mapping which questions competitors answer badly, checking terminology per canton: this used to be the part of an SEO project that quietly consumed the budget before anything shipped. AI turned days of it into hours. Deciding what was worth publishing, what the client could credibly claim and what was missing from the category stayed entirely with people.

The number that looked like a failure

Over six months, impressions went from 9.46k to 66k and clicks from 1.05k to 1.74k. Click-through rate went from 11.1% to 2.6%. Read as a single metric, that is a 76% collapse in CTR, and it is the sort of line that starts an uncomfortable meeting.

Here is what happened underneath it. At the start, nearly all impressions were brand searches: people typing the company name and clicking the first result, which is close to a guaranteed click. Six months later the site was appearing for thousands of non-brand queries at mid-page positions, where almost nobody clicks. Averaging a very small high-intent set with a very large discovery set pulls the percentage down by arithmetic, not by decline. Total clicks rose 66% in the same window, and average position improved from 16.4 to 10.3.

The structure that did the work

None of the AI leverage would have mattered without a plan for where the pages go. Three decisions carried most of the result.

Intent-mapped keyword clusters. Every target term was assigned to exactly one intent and exactly one page. Nothing competed with anything else on the site, which on a small domain matters more than it does on a large one, because there is no authority to waste.

Pillar pages that double as paid landing pages. The same page that ranks organically takes the Google Ads traffic. One build, two channels, and the paid data tells you within days which organic angle is worth expanding.

Bilateral internal linking. With no backlinks arriving from outside, internal links do that job. Every supporting article links up to its pillar and every pillar links back down, so authority circulates through the site instead of waiting to be donated.

The outcome is specific: Upgrid ranks second on Google Switzerland for strompreise schweiz gemeinden, above ElCom, the federal regulator whose data the page uses, and above ewz, 20 Minuten and the VSE. Only SRF sits higher. That is a new domain outranking national institutions on their own subject, with no link building at all. It is the same approach behind our SEO work generally, just under unusually tight constraints.

Getting cited by AI, not just ranked

Google AI Overviews and assistants like ChatGPT pick sources differently from a classic blue-link ranking, but the inputs overlap more than people expect. A model summarising a topic needs something it can quote without hedging: a number attached to a named source, a heading that matches the question, a table it can read cleanly, a definition that does not wander. Pages built to hold a reader for three minutes before revealing anything are exactly the wrong shape.

Upgrid now appears as a cited source inside Google's AI Overview for its category, and ChatGPT ranks it first among Swiss providers in its space, scoring it 9.8 out of 10. Neither result came from schema markup or a prompt trick. Both came from being the page that had the real figures, laid out so they could be lifted.

What to take from this

The useful question is not whether to use AI for SEO. It is which part of the job you hand over. Give it the building, the data plumbing and the research, and you get a team that spends its hours on expertise, judgement and genuinely useful pages. Give it the writing, and you get volume that costs nothing to produce and is worth roughly that.

Frequently asked questions