Keyword Density Died Quietly. 2026 SEO Is About Being the Source.

I still get briefs that say “keep the phrase at 1.5%.” I used to argue. Now I ask, gently, 1.5% of what — the page that nobody clicks becuase an overview alredy answered the query, or the page we want ChatGPT to cite on a Tuesday?

Counting how often a string appears was a workaround from a dumber index. It hung around becuase it is measurable, and marketing likes numbers that fit in a spreadsheet cell. In 2026 that cell is a souvenir. Search is a mix of classic rankings, AI Overviews on a huge slice of SERPs, standalone AI Mode that often ends with no click at all, and people asking Perplexity or Gemini insted of Google. Repeating “best running shoes Jakarta” seven times does not impress any of those systems. It just makes your writer tired.

This is not “SEO is dead.” That headline is also a souvenir. What died is the habit of treating a page like a bowl of keywords. What replaced it is messier, wich is why people keep reaching for density. Messy things need judgment. Density needed a plugin.

How we got stuck counting words

If you learned SEO from 2014 blog comments, density felt like physics. Too litle and Google “wouldn’t know.” Too much and you were a spammer. Tools put a score on it, bosses loved the score, writers stuffed. Google, for their part, spent years saying they moved on to entities, intent, links, and usefulness. We nodded and opened the density checker anyway.

Then generative content made the old trick industrial. You could ship 200 near-duplicates that all hit 1.2% of the same phrase. The response was not a new magic percentage. It was a wave of qaulity systems, spam policies, and answer engines that would rather quote one clear paragraph than scrape your seventh synonym.

By 2026 the public numbers look rude if you still think in “rank and collect ten blue links.” Zero-click behaviour is the majority of searches in some datasets. AI Overviews show up on a very large share of SERPs and, where they apear, organic CTR has been cut hard. AI Mode is even more zero-click. You can argue about the exact percentages you saw in the last conference slide. The direction is not subtle. Visibility is no longer only “position 3.” Visibility is “did the answer system use you, mention you, or skip you.”

What actually replaced the percentage

Not one metric. A pile. That is the point.

1. Intent fit, including the intent inside a prompt

People type like they talk to a model now. Longer, messier, “compare A and B for a family of four in Bandung who hates spicy.” A page that repeats a short keyword cannot cover that. A page that has a genuine comparison, cnostraints, and a next step can. Prompt-level research sits next to keyword research. If your tools only export head terms, you are planning for a SERP that is shrinking.

I keep a colum in briefs called “said like a human.” For a villa site in Canggu that might be “2BR with workspace and quiet at night, not a party villa.” That sentance will not have a pretty search volume. It will match how guests ask Gemini. Density cannot see it.

2. Entities and neighbourhood, not synonyms

Does the page mention the things a real expert would have to mention? For “kuliah kedokteran di Indonesia” that is UKMPPD, akreditasi, biaya kedokteran negeri vs swasta, internship, not twenty ways to say “medical school.” Coverage tools still help as a recap. They are a terrible religion. If a term cannot earn a spoken why, it does not go in.

3. Information gain (the unsexy phrase for “new stuff”)

If your article is a smoothie of the top 5, you are training data, not a source. Orignal photos of the proccess, a table you actualy compiled, a failure you had, prices you verified this month — that is what both classic ranking and citation engines have a cahnce to prefer. Indonesian exmaple: everyone has a “cara daftar NPWP” clone. The page that shows the current DJP Online UI, the 16-digit change, and what broke for a freelancer last week has gain. The page that hits 2% of “daftar NPWP” does not.

4. Evidence of experience, not adjectives about expertise

E-E-A-T is not a score you sprinkle. It is whether a human who did the thing left fingerprints. “We waited at KPP Setiabudi with nomor antrian X” beats “our team of experts.” Models can fake the second. They still struggle to fake the first if you require receipts — screenshots, dates, names of features that existed this quarter.

5. Structure that machines can lift

Clear definition in the first answer-shaped block. Headings that match questons. Tables. FAQ that is actualy asked. Schema that describes the page rather than a fantasy. This is not “keyword in H1.” This is “can an overview quote paragraph two withotu lying.” If paragraph two is stuffed, the quote looks stuffed, and you dont get picked.

6. Brand mentions and being known outside your own blog

Answer engines pull from a soup of the web, including forums and lists. Getting named in roundups, Reddit threads you did not fake, and local press is not “old school PR.” It is how you exist as an entity. A perfectly dense article on a brand-new domain is a whisper. I would rather see your UMKM cited in two honest roundups than see a 2% score in Surfer.

7. Technical hygiene as table stakes

Crawlable, fast enought, not a layout earthquake, indexation that matches the sitemap. Core Web Vitals still matter as experiance, and slow pages get skipped by impatient crawlers and humans. Density never loaded a hero image.

So… do keywords still matter?

Yes. Obviously. You still need to know what people type, which URL should own wich job, and wether two articles are punching each other. Keyword research did not die. Keyword density as a target died. The difference is the difference betwen knowing the city and counting how many times you say the city’s name in a postcard.

I still put a primary phrase in the title when it sounds like a title a person would click. I still use it once early, if it fits. I do not keep a running percentage in a plugin while I write. If the outline is right, the phrase apears as often as a normal explanation needs, wich is almost always less than the plugin wanted and more than zero.

For bilingual Indonesian sites this is extra importent. Forcing an English keyword percentage into a Bahasa article makes the prose feel translated, and both Google and readers notice. Pick the langauge of the query. Use the other langauge as a clrifying alias, once, if people mix terms (NPWP / tax ID). That is entity hygiene, not stuffing.

What I measure instead on Monday morning

If you need a dashboard so you dont feel unmoored, try this set. It is not perfect. It is more honest than a density colum.

• Query → URL ownership: one primary job per URL. Cannibalisation list, expecially Bahasa/English twins.
• Impression/CTR splits where overviews likely apear: position held, CTR fell. That is a snippet and uniqueness problem, not a “add keyword” problem.
• Cited or not: a lightweight check — do you show up in a handful of tracked prompts in ChatGPT / Perplexity / AI Overviews for your money queries? Even a manual monthly pass beats never looking.
• Unique assets per URL: a checkbox. Table we built, photos we took, numbers we verified. If the box is empty, do not publish another 2,000 words.
• Engagement that is not vanity: scroll to the tool, click the calculator, time on the comparison table. Bounce on a dictionary-definition page that alredy lost to the overview is not a writting failure. It might be a “this URL should not try to rank for that query” failure.

Agencies hate that this is not one score. Founders like one score. You can give them “number of money prompts we are cited on” if they need a single number. It will change slower than density and mean more.

A page I would actually ship in 2026

Topic: buying a used Avanza in Jakarta, a query that is half dream and half scam-avoidance. Old SEO page: 1,800 words, “beli Avanza bekas” every 80 words, generic tips, stock photos of a diffrent car.

Page I want:

• A blunt 80-word answer: what year/range still makes sense, what to inspect, when to walk away.
• A table of common failure points with costs in rupiah you checked with a bengkel this quarter.
• Photos from a real inspection — odometer, chassis number, flood line — ugly and usefull.
• A section on OLX / Facebook Marketplace red flags that are local, not Craigslist folklore.
• Internal links to financing and to a seperate “hindari mobil bekas banjir” URL, not a mega-essay.
• Author line with the mechanic’s name, not “Staff Editorial.”

Count the primary phrase if you must. It will land where a human would put it. The reason it can rank or get cited is the table and the photos and the fact that a preson put their name on the flood line. Density did not do that work. It never did. We just pretended.

How AI should sit in this workflow (without becoming the product)

Use models to cluster queries by job, to diff your outline against the live SERP, to flag entities you forgot, to draft meta in five variants you will rewrite. Do not use them to hit a percentage. Do not use them to generate 40 pages a week that all “cover” a topic with the same spine.

The sites getting hurt are not the ones that used AI as a pair of hands. They are the ones that used AI as a factory for pages that never had a reason to exist except a keyword gap spreadsheet. Quality systems and answer engines are aligned on this, mayb for the first time: unoriginal bulk is expensive to host and easy to skip.

If your content calender is a list of head terms, change the calender object. Each row should be a queston a buyer asked this month, a proof you can add, and a URL that alredy exists or a reason it must exist. If you cannot name the proof, you are about to write density again, just with nicer transitions.

What to tell your boss, in one paragraph

We are not abandoning keywords. We are abandoning the idea that repeating them is a ranking strategy. In 2026 people see answers withotu clicking, and the systems that write those answers look for clear, specfic, citable pages from sources the web alredy trusts. Our job is to own a query with one URL, put somehting on it you cannot copy-paste from the top 5, and make it easy to quote. The plugin that scores 1.5% cannot help with that. A mechanic with a camera can.

If they still want a number, give them citations and conversions, not a percentage of a string. It feels less scientific for a week. Then a page you actualy cared about starts showing up in the overview, and the conversation ends.

Your next three moves

1. Delete density as an acceptance criterion in your brief template. Replace it with “unique asset” and “queston answered in 80 words near the top.”
2. Pick ten money prompts your buyers would type into a chatbot. Check who gets cited. That list is your real competitive set, and it may not match the classic top 10.
3. Refresh one existing URL with proof (new table, new photos, new prices) insted of publishing three new essays that restate the SERP.

Keyword density had a good run as training wheels. Taking the wheels off feels wobbly. It is also how you stop competing with a million pages that all learned the same trick from the same model.

Write like you have to put your name on it in a WhatsApp group of people who know the topic. That audience was always the ranking system. We just built plugins so we could forget.