Skip to content
notes

Where the contested questions get settled.

Most writing about AI visibility is a thin definition wrapped around a lead form, or a confident take that the data does not support. Notes is the opposite: dated, sourced dispatches on the questions the field argues about — and gets wrong — written the way we tell clients to write, so the engines that we cover can cite them. No gate, no fluff, the contested parts flagged as contested.

latest

Published

The organic window is not being replaced — it is splitting: AI ads and citation

The panic headline says AI ads are closing the organic window. 2026 says otherwise: OpenAI put ads below the answer, Perplexity pulled them, Anthropic refused, Google integrated them. Ads and citation are separate lanes — and earned authority is the bet you never re-place.

Read →

Measuring AI visibility without fooling yourself: leading indicators and the limit of tools

AI citation is non-deterministic and the tools sample an infinite prompt space, so most dashboards measure noise. The honest way: trend over level, crawl logs as a leading indicator, referrals as confirmation.

Read →

SEO is not being replaced — it is decoupling: the 70%-to-under-20% story

The overlap between what ranks in Google and what AI engines cite has fallen from ~70% to under 20%. SEO is not dying; it is decoupling into a separate discipline. What that means for where you spend.

Read →

What llms.txt actually does in 2026 (and what it doesn’t)

One camp sells llms.txt as a ranking factor; the other calls it snake oil. The 2026 data supports neither. The honest verdict — not a search-citation lever, but a real agent-to-agent signal — with the sources.

Read →

Why AI doesn’t cite you: the three stages where a citation breaks

Most GEO advice tells you how to get cited. This is the opposite: a diagnostic map of why you don’t — parsing, fetching, generation — built on the first systematic taxonomy of citation-failure modes.

Read →

The five-domain myth: concentration and the long tail in AI citations

“Reddit gets 40% of citations” and “no domain exceeds 5%” are both true, from different 2026 datasets. How concentration and the long tail coexist — and what it means for a brand that isn’t a giant platform.

Read →

The Great Decoupling: your rankings held, your clicks left

Impressions at record highs, clicks falling, rankings untouched. The Great Decoupling explained: why it happens, what the February numbers show by vertical, and what to measure when the click can no longer be trusted on its own.

Read →

Google AI Mode: one question becomes sixteen, and you are cited or invisible

AI Mode runs on Gemini 3 as of January 2026 and shows no blue links: you are cited or invisible. It works by query fan-out — one question split into a dozen-plus concurrent sub-searches. How it works, and how a deep page wins sub-questions it never targeted.

Read →

Block the AI crawlers and you vanish from the answers they feed

Millions of sites blocked AI bots in 2025 — but "block the AI crawlers" hides a distinction. GPTBot trains; OAI-SearchBot and PerplexityBot cite you. Block the wrong one and you delete yourself from the answers. The training-vs-retrieval call, bot by bot.

Read →

For AI, being mentioned beats being linked: the 0.664 vs 0.218 gap

A 75,000-brand study found web mentions correlate with AI visibility at 0.664 and backlinks at just 0.218 — roughly 3x. AI reads your brand as an entity built from language, not the link graph. Why earned mentions beat links, and where to move the budget.

Read →

There is no "optimize for AI": each engine cites a different web

AI search is not one engine. A 6.8M-citation analysis found Gemini favors your own structured site (~52% brand-owned), ChatGPT leans on directories and references, Perplexity on industry directories and community. Low overlap — so there is no single AI playbook.

Read →

AI quotes the forum, not your blog: why UGC is the most-cited source

In 2025 the most-cited AI source is not a brand blog — a 150,000-citation study put Reddit first at ~40%, ahead of Wikipedia, YouTube and Google. AI rewards the human signal of lived experience. Why community wins, the honest catch, and what to actually do.

Read →

Schema and AI citations: useful, oversold, and not what they sold you

Microsoft says schema helps its LLMs; Google says it is not a ranking factor and calls it a stopgap; a 2024 study found no schema–citation correlation. What structured data really does (disambiguate your entity) versus what it is sold as (a citation lever), and how to do it without overpaying.

Read →

AI doesn't read your page — it lifts a chunk. Write for that.

AI does not read your page top to bottom — it chunks it into passages and cites the ones that stand alone. A 2025 study found 40-75 word answers cited ~3x more. Answer-first writing, question-style H2s, and why ranking is not enough: the fix is usually structural, not authority.

Read →

Getting cited is not the goal if the citation is false

AI hallucinations attribute false claims to your brand, invent executive quotes, and misstate your product — in the same confident tone as the truth. 2024 studies found 58-88% on legal questions, and some newer models hallucinate more, not less. Measure accuracy, not just presence.

Read →

People stopped typing keywords and started asking questions

With AI Mode rolling out in May 2025, search shifts from telegraphic two-word keywords to full natural-language questions that pack intent, context and constraint into one query. Why this rewards being the authoritative answer to a decision over targeting a string.

Read →

The one thing AI cannot synthesise from somewhere else

Generative engines synthesise what already exists — they cannot produce primary data. Publish original numbers only you have (product telemetry, surveys, category benchmarks) and an answer that needs that fact has nowhere to go but you. Substance, not just format.

Read →

The industry is naming the shift faster than the shift is changing

SEO, AEO, GEO, LLMO, AIO — the field is minting acronyms faster than the practice changes. What each actually means, where they overlap (most of it), and why the label matters far less than the shift from ranking links to being the source of an answer.

Read →

When AI stops answering and starts researching

In February 2025 a new kind of AI reader arrived: the deep research agent (OpenAI on Feb 3; Google had one in Gemini since December 2024). Unlike a chat that cites a handful of sources, it reads dozens or hundreds and writes a cited report — rewarding depth and consistency over surface tricks.

Read →

Stop asking what your position is in ChatGPT

AI responses are non-deterministic by design — the same prompt gives different answers each run — so there is no rank to track. Visibility is a probability of appearing, measured by sampling a prompt many times, not a fixed position. Stop chasing a number the system does not produce.

Read →

Two ways to show up in AI: remembered, or retrieved

On 16 December 2024 ChatGPT Search opened to all logged-in users — the most-used chatbot began searching the web and citing sources. There are now two ways to appear in AI: remembered, from training, or retrieved, live, with a clickable link you can measure.

Read →

When the generated answer reached a billion searchers

In late October 2024 Google rolled AI Overviews out to 100+ countries, Spanish included — a generated answer with citations, above the links, for over a billion searchers. Being one of the cited sources inside the answer is a new kind of visibility, on a different surface from the chatbots.

Read →

Why earning mentions matters more than polishing your own pages

Audits in 2024 show AI search leans on third parties — press, reference sites, communities — and rarely a brand’s own pages. So appearing in AI depends heavily on earned media: being mentioned by the sources the model already trusts. Digital PR, not on-site polish, is the lever.

Read →

AI is shaping your buyers, and your analytics can’t see it

In 2024 LLMs became private research advisors for B2B buyers — and a chatbot conversation leaves no click, cookie or referral. The moment AI shapes a buyer is structurally invisible; when they arrive later it is mislabelled as direct or brand search. You measure this channel by proxy, or you value it at a false zero.

Read →

The study that named GEO, and what it proved works

At KDD 2024, the Princeton-led paper that coined Generative Engine Optimization gave the first controlled evidence of what works: adding statistics, citations and quotations lifts a source’s visibility by up to ~40%, while keyword stuffing drops below baseline — and the best method depends on the domain.

Read →

The entity problem: recognised before cited

Engines reason about entities, not strings. Before an AI can cite your brand, it has to recognise you as a specific, disambiguated entity — in the knowledge graph, described the same way everywhere. The entity layer is the invisible foundation of AI visibility: recognition comes before citation.

Read →

When the answer engine told people to eat rocks

In May 2024 Google’s AI Overviews told people to put glue on pizza and eat a rock a day — pulling a sarcastic Reddit joke and a satirical article in as fact. Hallucination is the nature of these systems, not a bug to be patched, and the quality of sources on your topic decides whether AI gets it right.

Read →

Your site says you’re good; reviews prove it — and AI trusts the proof

When a B2B buyer — or an AI answering for them — asks "what’s the best X for Y?", the answer comes from independent reviews and comparison platforms (G2, Capterra, TrustRadius), not your own site. Your site is an interested party; a peer review is evidence — and AI inherits that earned authority.

Read →

Generative search has been telegraphing itself for a year

Generative search did not arrive overnight. Since May 2023, Google has tested it in public as SGE — a synthesized answer with cited sources, above the blue links. As of April 2024 it is still experimental, but the direction is unmistakable, and preparing to be a source it can quote is an early advantage.

Read →

Volume is not value: what the March 2024 update measures

On 5 March 2024 Google made "scaled content abuse" a spam policy — mass-produced pages built to manipulate rankings, whether written by AI, humans, or both. The lesson for AI visibility is one of measurement: count whether a page answers better than the alternatives, not how many pages you shipped.

Read →

Parametric vs retrieval: where an AI knows you from

Google renamed Bard to Gemini this month and the debate was which model is smarter. For your brand, the question is which of an AI’s two memories the answer came from: the trained-in half you cannot edit, or the retrieved half you can.

Read →

Corroboration is the authority an answer engine trusts

A young answer engine raised at around a half-billion-dollar valuation this month on one promise: answer, and cite the sources. The citation reveals what authority an engine rewards — corroboration across credible sources, not one loud mention.

Read →

Optimize for the question, not the model

Google launched Gemini this month and the leaderboard moved again. Chasing the model of the moment is a losing strategy — models churn every few months; the question your buyer asks does not. Build for the question.

Read →

What survives when the platform shakes

For five days this month the company behind ChatGPT nearly came apart. The reminder for brands: your position inside a platform is rented — the authority you earn off it is the part you own, and the part that survives.

Read →

The engines learned to see

This month ChatGPT gained eyes — it can read the text inside an image. A layer of your work just stopped being invisible to machines: put real text in your pictures and keep your entity consistent across every mode.

Read →

The answer won’t hold still

Bard began tailoring answers to your own Gmail and Docs. A generative answer is not one fixed thing you can look up: stop snapshotting a single response and start sampling the distribution — count how often you appear, not whether you appeared once.

Read →

Your buyers brought AI to work

OpenAI gave companies a private, sanctioned ChatGPT. The barrier to AI at work was never interest — it was data risk — and removing it moves AI earlier in how your buyers research and shortlist. You now have a reader to satisfy before the human ever sees you.

Read →

Your authority doesn’t translate itself

Bard reached the EU and Brazil in 40+ languages. As the answer engines go multilingual, the reputation a model recognizes is built language by language — being well known in English does not make you known in Spanish. Earn depth where your growth buyers ask.

Read →

The model knows an old version of you

OpenAI shipped newer models with function calling — and a knowledge cutoff still stuck in September 2021. A newer model is not newer knowledge: what a model memorized about your company can be years out of date. Measure what it says and keep your record straight.

Read →

The snapshot quotes a sentence, not a page

At I/O, Google previewed a search that synthesizes one answer from several sources. The unit of use shrinks from the page to the sentence: a claim has to be self-contained and attributable to survive being extracted and recombined — or be misread, miscredited, or dropped.

Read →

When the AI started acting, not just answering

Auto-GPT and BabyAGI went viral — experiments that let a model chase a goal on its own. They barely work: loops, hallucinated steps, burned budgets. But the direction is real, and the sober way to prepare is not a new tactic, it is the old hygiene: clear, verifiable, corroborated information.

Read →

Ask the model what it says about you

GPT-4 landed and a capable model now sits in everyone’s browser tab. The cheapest first move on AI visibility: ask it about your own company and read the answer in four buckets — accurate, stale, invented, missing. A diagnosis, not a metric.

Read →

The model describes you in words you didn’t choose

The new Bing turned search into a conversation — and the shop window stopped being yours. A blue link showed the title and description you wrote; an AI answer paraphrases you in its own words and tone. You control the raw material, not the sentence.

Read →

The model remembers you, or it doesn’t

ChatGPT became the fastest-growing consumer app ever — but it answers from memory, with no search, links or sources. There is no index to climb and no tag to tune. The model learned your entity from the public record, or it did not.

Read →

The day search became an answer

A free chatbot went from launch to a million users in five days, and people started asking for an answer instead of a list. A list lets you compete for a spot; an answer includes you or it does not — and can include you wrongly. Where this archive begins.

Read →

Position zero was already an answer

Long before any chatbot, search was already an answer: featured snippets, People Also Ask, position zero — but it cited its source. As November ends, two signals of where it accelerates: a science model pulled for confident falsehoods, and a research-preview chatbot on the 30th.

Read →

Authority isn’t a tag you add

The Helpful Content Update and this month’s spam update point the same way: search tells people-first pages from pages written for it. You cannot add E-A-T to a page — it is a reputation earned off the page, and that is the only kind that lasts.

Read →

Don’t make the engine guess

Without structured data, the engine infers what your page is — and inference can be wrong. Schema lets you state the facts instead. It is not a ranking factor and never was; it is legibility, the difference between being read correctly and being guessed at.

Read →

Would you write this if search didn’t exist?

The Helpful Content Update, rolling out now, turns a question that used to be philosophical into a ranking signal: are you writing for people or for the engine? It distils to one test — would you make this if search engines did not exist?

Read →

The review you didn’t test shows

Google’s July product reviews update — the fourth in fifteen months — keeps rewarding evidence you actually used the product: original photos, real measurements, what the spec sheet left out. The engine is learning to measure the difference. The one you didn’t test shows.

Read →

A core update isn’t a penalty

The May core update finished this week, and the instinct is to ask what you did wrong. But a core update is not a penalty — it is a re-evaluation of relative quality, and dropped pages have nothing to fix. The only response that works is becoming a more authoritative answer, especially in YMYL.

Read →

The engine might think you’re someone else

If another organisation shares your name, the Knowledge Graph can confuse the two of you — blending your signals, mixing your knowledge panel, crediting your authority to your namesake. Disambiguation is how you tell the engine which entity you are.

Read →

Core Web Vitals won’t save weak content

Page experience reached desktop last month, and the temptation is to treat Core Web Vitals as a ranking lever. They are not — Google is explicit that page experience is a tiebreaker, not a substitute for great content. A fast page with weak content still loses.

Read →

The quality raters don’t grade your site

Google employs thousands of human quality raters — but their scores do not touch your ranking. They calibrate the algorithm in aggregate; no rater grades your page. Read the guidelines as the clearest public picture of what Google means by quality.

Read →

A hundred links don’t beat one that matters

The instinct is to pile up backlinks. Since Penguin, Google weighs quality, relevance, and naturalness, not the count — one earned editorial link beats a hundred low-quality ones, which can actively hurt you. Authority from links is earned, not manufactured.

Read →

Schema wins you the look, not the rank

The hope is that structured data lifts your position. It does not — schema is not a direct ranking factor. What it wins you is eligibility for rich results: stars, FAQs, breadcrumbs, prices that change how your result looks and can raise its CTR at the same position.

Read →

Your reputation is written where you can’t edit it

The trust in E-A-T is the one part of quality you cannot put on your own site. Google’s quality raters judge reputation from independent sources — third-party reviews, mentions, coverage — not your About page. When your claims conflict with the outside record, Google believes the outside.

Read →

Average position is a story, not a number

Search Console’s average position is one impression-weighted figure across every query you appear for. It can rise while your valuable queries fall, and fall while nothing changed. Read alone — especially during a core update — it invents a story your traffic does not support.

Read →

The title in the results isn’t always the one you wrote

Since late August 2021, Google rewrites the title link shown in results — pulling from your H1, other headings, prominent text, even anchor text — instead of always using your title tag. It is

Read →

Core Web Vitals are a tiebreaker, not the prize

Google confirmed the page experience update finished rolling out in September 2021, making Core Web Vitals — LCP, FID, CLS — an official ranking factor on mobile. But it is a small one: a

Read →

Links you didn’t earn can be switched off

Google’s link spam update finished rolling out in late August 2021. The thing to understand is the mechanism: it does not punish your site for bad links — it nullifies them, switching off the

Read →

A core update isn’t a verdict on your page

The July 2021 core update finished rolling out on 12 July, the second half of a core update Google split across June and July. The thing to understand is what it is and is

Read →

Which update hit you? The fix depends on the answer

June 2021 was loud: a core update from 2–12 June and two spam updates on 23 and 28 June, constantly confused for one another. They are opposite kinds of event. A spam update is

Read →

Write for an engine that understands, not one that matches

At I/O this month Google announced MUM, a language model it says is far more capable than BERT — multimodal, trained across many languages, built to handle complex questions. It is not deployed and

Read →

A good review earns its rank; a summary borrows it

Google’s first product reviews update finished rolling out this month. It rewards reviews that share insightful analysis and original research from someone who actually knows the product, and lets thin content that merely summarises

Read →

Don’t make Google dig for your answer

Google’s passage ranking went live last month: it can now rank a single buried section of a long page for a query, even when the page as a whole is about something broader. Google

Read →

What Google says about you, before the click

Google’s new About this result panel went live this month. Tap the three dots beside a result and Google tells the searcher who the source is — a Wikipedia description if one exists, otherwise

Read →

Most of the movement is noise, not signal

A core update finished rolling out last month, and now the dashboards are full of motion: rankings up, rankings down, every wobble inviting a story. But Google itself says positions are not fixed —

Read →

If it isn’t on mobile, it isn’t indexed

Google has set the end of March 2021 as the point when every remaining site moves to mobile-first indexing. From then on it reads, indexes, and ranks your site from the mobile version —

Read →

Six months’ notice is the work, not the wait

Google almost never tells you a ranking change is coming before it arrives. This month it did: the page experience signal will become a ranking factor in May 2021, about six months out. That

Read →

Cover everything, own nothing

At its Search On event this month, Google described how it now uses neural nets to understand the subtopics inside a broad interest — so a search for something general can return a spread

Read →

Google indexes the document, not the design

Google quietly clarified this month that it may not index text placed in CSS content properties, because that text is not part of the DOM. It is a small note with a large principle

Read →

Your speed shows before they click

Google announced this month that Chrome for Android will start labelling links as a “Fast page” — a badge shown in the context menu, before anyone visits, for pages that have historically met the

Read →

What your query report leaves out

Search Console has a quiet limitation worth understanding before you trust its numbers: it does not show you every query. Rare queries — the long-tail ones searched by very few people, and any that

Read →

When Google can’t tell it’s all you

Google does not just index pages; it builds a model of the entities behind them — your company, the people in it, the things you make — by gathering signals from across the web.

Read →

Experience is a tiebreaker, not a substitute

Google announced this month that page experience — a new signal combining Core Web Vitals with existing UX factors like mobile-friendliness and HTTPS — will become a ranking factor next year. The detail that

Read →

When the dip is the dashboard, not your site

Early this month Google confirmed a bug that de-indexed some pages and then disrupted Search Console’s reporting — coverage numbers fell, the URL Inspector lagged, dashboards showed drops. The pages were fine; the data

Read →

Your link tags are hints now, not commands

As of the first of this month, the nofollow attribute stopped being a directive Google obeys and became a hint it weighs — joining the sponsored and ugc attributes announced last September. The rel

Read →

What Google won’t confirm, it can reverse

Over the weekend of February 7–9 rankings shook hard — sites gained or lost top positions across every kind of niche — and by the 13th Google had quietly rolled almost all of it

Read →

You earn the snippet, not a second listing

On January 22 Google deduplicated featured snippets: a page elevated into the snippet no longer also appears in the ten blue links below it. What used to be “position zero” — a bonus box

Read →

Context decides what your words mean

On October 25 Google announced BERT, which it calls its biggest leap in five years — a model that reads a query in both directions at once, so the words before and after a

Read →

You get two slots, not the whole page

In early June Google rolled out a site diversity change: for most queries it will no longer show more than two listings from the same domain in the top results. The measured impact was

Read →

The higher the stakes, the higher the bar

On August 1 Google rolled out a broad core update — the SEO community named it Medic — that hit health, medical, and financial sites hardest. A week earlier Google had updated its Quality

Read →

Your thin pages drag down your good ones

In late February Google launched the update the industry first called Farmer and now calls Panda — a change that hit content farms and thin, low-value pages hard, touching nearly 12% of queries. The

Read →

Being in the index is the starting line

In June Google announced Caffeine, a rebuilt indexing system that replaces the old layered, batch-refreshed index with a continuous, incremental one — about 50% fresher, and able to index a page soon after you

Read →
why notes exists

The proof, not the pitch

The whole argument we make to clients is that AI engines cite content that answers a question directly, backs it with specifics, and comes from a credible, consistent source. A blog that hid its substance behind a form, or asserted things it could not source, would contradict the method on its own page. So Notes is written the way we would write for a client who wanted to be cited: answer-first, evidenced, and honest about what is genuinely unsettled.

That last part is the point of difference. The field is full of confident claims in both directions — llms.txt is everything, llms.txt is nothing; SEO is dead, GEO is just SEO. The useful truth is almost always in the middle, in the data, with the contested parts named as contested. The AC Group has earned attention online for 27 years by being the substantive source rather than the loudest one; Notes is that discipline applied to the questions this field keeps getting wrong.

coming next

On the desk

The next dispatches, each aimed at a question the field answers badly. These are in progress, not yet linked:

Reading is the first step. The audit is the second.

When you want to stop reading about AI visibility and see your own, the free snapshot shows how five engines name you today and what would move first. Forty-eight hours, no sales call.