Co-occurrence

What co-occurrence actually measures

Two units turn up inside the same window of text, repeatedly, across a corpus large enough to rule out chance. That is the entire mechanism. Everything that matters sits in the two parameters most practitioners never state: the window, meaning how much text counts as « together » (a sentence, a paragraph, five tokens either side of the pivot, a whole document), and the corpus, meaning what body of texts you measured on. Move either one and the same pair of mots goes from significant association to statistical noise. When someone tells you a term co-occurs with your brand and cannot tell you the window, they are describing an impression, not a measurement.

The definition proper, since it is useful embedded rather than up front: co-occurrence is the simultaneous presence of two lexical units within a delimited stretch of discourse, counted across a corpus. It is a raw observation about textual distribution, not a claim about meaning. The interpretive step comes later, when you decide whether the frequency of the pair exceeds what independence would predict.

On the spelling, since it costs the sector more attention than it deserves: English usage keeps the hyphen after the co- prefix before a doubled vowel, French dictionaries list « cooccurrence » welded in one word, and both refer to the same notion. No analyzer worth using treats them as distinct: normalisation folds them into one token before any statistic is computed. Pick one form, stay consistent inside a document for readability, and move on. There is no retrieval consequence.

Collocation, corpus and the linguistic root

The idea predates search engines by decades. Firth's 1957 formulation, that you shall know a word by the company it keeps, is the founding line of distributional semantics, and every vector model shipped since runs on some refinement of it. Corpus linguistics and French lexicométrie built the measurement apparatus: frequency tables, contingency matrices, association scores computed over annotated textual corpora rather than intuition.

The distinction that keeps getting flattened in SEO writing is co-occurrence versus collocation. Co-occurrence is the observation: these two unités appeared near each other, n times. Collocation is the conclusion: the pair recurs far above chance and has become conventionalised, so much so that a native speaker anticipates the second word from the first. « Heavy rain » is a collocation. « Rain » next to « Tuesday » in a weather corpus is co-occurrence and nothing more. Every collocation is a co-occurrence; the reverse is almost never true, and the gap between the two is exactly what the statistical measures in section five are for.

Why a netlinking professional should care about the vocabulary of a dictionnaire rather than treating it as academic decoration: the maths under semantic SEO tooling is the same maths. Content optimisation scores, entity extraction, embedding similarity, topic clustering, all of it descends from counting co-occurrences in a corpus and normalising the counts. Understanding the primitive tells you which vendor claims are measurable and which are dressing.

Why it became a citation signal in 2026

For fifteen years the profession optimised the anchor, because the anchor was the cheapest place to state a topical claim to Google. That arbitrage is closing. Google's spam update of 26 August to 22 September 2025 hardened enforcement against manipulative anchors, and Google's own wording is that once spammy links are neutralised, the ranking benefit cannot be regained. Analyses of the 2025 Link Spam Update converge on the same reading: the system now weighs the context, the relevance and the relationship between domains, not the anchor string in isolation. A March 2026 review of Google's backlink policy by Bluetree, looking at 85 B2B SaaS domains, found the ones that held or improved rankings carried at least 38% branded anchors, which is another way of saying their link profiles read as mentions rather than as placements.

Then generative retrieval arrived and made the shift structural. Ahrefs' correlation study across 75,000 brands, published 27 April 2026, ranks brand web mentions first at r = 0.664 against presence in Google's generated answer panels, ahead of brand-rich anchors at 0.527 and brand search volume at 0.392. Ahrefs states the mechanism plainly: language models derive authority from the prevalence of particular words, the co-occurrence of different terms and topics, and the context those words sit in. Search Engine Land put a name on it on 4 May 2026, « brand co-occurrence », and argued it is a stronger lever for AI search visibility than topical authority considered on its own.

The decoupling is measurable. Surfer's analysis of 36 million AI Overviews and 46 million citations found only around 17% of cited sources also rank in the organic top 10 for the same query. Ranking and being cited are now two markets with two entry conditions, and the second one runs on association rather than position. If you are building for visibility inside generative engines, the unit of work stops being a ranking URL and becomes a pattern of textual proximity between your brand name and your category terms, repeated across enough independent documents to survive as a statistical regularity. That is what our work on getting a brand named inside AI-generated answers actually optimises.

Where it fits in a netlinking operation

Semrush's July 2026 guide on what works in link building draws the line we use internally. Co-citation is your brand being named or linked alongside other trusted brands, the way appearing in the same comparison list as an established player positions you by adjacency. Co-occurrence is the terms related to your business recurring near your brand name and your links. Semrush's point, and it matches what we see, is that editorial quotes and directory entries can create both without moving meaningful link equity, and still change how systems categorise the brand.

Operationally this changes the brief, not the budget. The anchor is one phrase; the paragraph around it is 60 to 100 words of free signal, and most campaigns waste it. A placement whose surrounding text never names the category, never names a competitor, never states what the brand does, delivers a link and no association. We run 50 owned French editorial media in-house, and the brief we hand our writers specifies the semantic field the paragraph has to occupy, not just the anchor and the target URL, because the paragraph is the part that survives an anchor devaluation.

The practical consequences: prefer placements where the article topic genuinely contains your category vocabulary over higher-authority pages on unrelated subjects, since a link from a strong domain in the wrong lexical field produces no useful co-occurrence. Accept unlinked mentions on relevant pages rather than discarding them as worthless, because the Ahrefs correlation puts them at the top of the table. And treat the mix as something to calibrate over several months rather than buy in one batch, because association is a frequency effect and a single placement cannot produce one. When you need to see which médias actually publish in your semantic field before committing, the public catalogue of sites with their real topical coverage is the faster filter.

Measuring it: PMI, log-likelihood and matrices

Start with the contingency table: how often does term A appear, how often B, how often both inside the chosen window, and how large is the corpus. From those four numbers every association measure follows. Pointwise mutual information compares observed joint probability against the product of the marginals, so it answers « how much more often than chance ». Log-likelihood ratio tests the same hypothesis but stays reliable on low counts, which is why corpus linguists reach for it first. Chi-square works on dense data and degrades badly on the sparse tails where most interesting pairs live.

The construct behind all of this is the co-occurrence matrix: rows and columns of vocabulary items, cells holding joint counts inside the window. This short explainer covers how that matrix is built and where it sits among the other classic text-feature representations:

For the mathematical treatment, including how the matrix behaves as dimensionality grows and why it gets factorised in practice, this lecture goes a level deeper:

Thresholds that hold up in practice: a symmetric window of five tokens either side, clipped at sentence boundaries, for lexical association; document-level windows for brand and entity association, since a brand named in paragraph two and its category named in paragraph nine belong to the same claim. Filter on a minimum joint frequency of five before reading any PMI ranking, otherwise the top of your table is hapax pairs with perfect scores and zero meaning. On tooling, AntConc handles concordances and collocates without writing code, TXM covers the lexicométrie side including textual statistics over structured corpora, and a Python stack gives you the matrix directly when you need to run it over a scrape of your own SERP set. On the SEO side, the corpus you actually want is the set of pages that rank or get cited for your category terms, and the question is whether your brand appears in it at all.

What we see go wrong

The dominant failure is treating co-occurrence as keyword density with a new label. Repeating your brand name beside a category term twelve times on your own page changes nothing: the signal is cross-document, computed over a corpus you do not own. Your site is one point in that corpus and its self-declarations are the least trusted part of it.

The second failure is buying volume on pages nobody reads and nothing cites. Co-occurrence computed over the indexed web weights sources unevenly, and a mention on a page with no citations of its own contributes close to nothing. The third is ignoring the sign of the association. Association measures are blind to sentiment, so a brand consistently occurring near « alternative to » or near complaint vocabulary is building an association it did not intend, and it is the kind of pattern that turns up in generated answers before anyone thinks to audit it.

The last one is impatience. A single quarter of placements moves a co-occurrence profile by an amount indistinguishable from noise on any corpus large enough to matter. Measure it quarterly as a share of your category's document set, not weekly as a count, and accept that this is the slowest-moving lever in the toolbox and also the one that anchor-level enforcement cannot take away from you.