Keywords and cannibalization
How a page's primary keyword is inferred from its own content, what the three conflict sections mean, why translations and shared vocabulary are not conflicts, and what blocking a term does.
The Keywords page answers two questions per page: what is this page about, and is another page on the site about the same thing. Both answers are read out of your own crawled content. Search Console, when connected, adds a third source of evidence, but nothing here needs it.

What the page reads
Everything comes from the last crawl of the page itself: the title, the H1, the URL, the meta description, the first paragraph of the body, breadcrumbs, Open Graph tags, structured data, H2s and H3s, image alt text and meta keywords. That is thirteen signals, and they are not equal: the title counts most and meta keywords least.
The page body has to have been read, which every crawl now does; a page marked Meta only on the Content page was crawled before that and needs a re-crawl. A site with no crawl at all has nothing here, and the empty state says so.
Primary and secondary
Every phrase of one to five words that appears in a signal is scored: each signal it appears in adds that signal's weight, and a phrase several independent signals agree on scores far higher than one that appears in only the body. An exact match with the title or the H1 adds more. The best-scoring phrase is the page's primary keyword; the next five are its secondary ones.
The dots beside a phrase show which of the thirteen signals it was found in, darker for the heavier signals. More dots means more confidence.
Once the whole site is scored, a second pass adjusts for the site itself: a phrase that appears on nearly every page is pushed down, and a word that frames almost every phrase, the company name for instance, is trimmed off the edges. That pass needs at least ten scored pages, so on a very small site the raw scores stand. It is also why a page's primary can change after a re-run even though the page did not.
The three kinds of conflict
The Conflicts tab has three sections, in order of how much evidence sits behind each.

Measured in Search Console is what Google reports: two or more of your pages appeared for the same query in the top twenty positions over the last 28 days. The page with the most impressions is the winner, which is not always the best-ranked one, and the severity comes from how many impressions went to the other pages. Search Console only keeps each page's top fifty queries, so this section shows conflicts it can prove, not all of them.
Confirmed is inferred from content: two or more pages hold the same phrase as their primary keyword. They are trying to be about the same thing.
Worth a look is the weaker shape: one page owns the phrase as its primary and one or more others mention it among their secondaries. Single-word phrases never land here, because one word shared by a few pages is vocabulary, not competition.
Two rules limit the list. Conflicts are grouped within one language, so a page and its translation are never reported as competing. And a phrase that many pages merely mention, with none holding it as primary, is not a conflict at all.
Your own name is handled separately. Queries that are your brand are counted in a footnote and given their own filter, with a neutral badge, and they never appear in the headline count. A site ranking for its own name with several pages is expected.
Two counts that differ
The headline counts conflict groups: one phrase, several pages. The severity filter on the All pages tab counts pages. A phrase five pages target is one conflict in the headline and five rows in the table.
The All pages tab
One row per page with keywords: its primary, a score bar, and a badge for the worst conflict it is in. The badge names the conflicting phrase, which may not be the page's own primary, and says whether the conflict was measured or inferred; when both exist, the measured one wins. Filter by High, Medium, Low, Your own name or Conflict-free, and search by URL, title or phrase.

Expanding a row shows the secondaries with their signal dots, and each conflict in full: the competing pages, their scores or impressions, and the winner.
Low is a severity only Search Console produces. Inferred conflicts are High or Medium, so on a site without Search Console the Low filter is empty, and that is expected.
Blocked and frequent terms
A blocked term is removed from every phrase before scoring. Use it for brand names and boilerplate that turn up on every page and would otherwise become primaries. Frequent terms lists the phrases found on the most pages, with a Block button beside each, which does the same thing.

Blocking a term changes nothing until the next analysis. Press Re-run after editing the list.
Re-run
Re-run scores any page that has not been scored yet, then runs the site-wide pass over every page and rebuilds the conflict list. It runs on its own after every crawl and on a schedule as a safety net, so the button is for after you change blocked terms or want the result now. Running it twice in a row changes nothing the second time.
The graph
The Keywords view on the Visualize page draws the inferred conflicts as clusters: a phrase in the middle, the pages that target or mention it around it, coloured by category. Search Console conflicts are not drawn there. Low groups are hidden by default, and Min URLs narrows the graph to phrases shared by at least that many pages.
Deciding what to do
Pick the page that should win. For a measured conflict, start from the page Google already prefers. If that is the wrong page, change the titles first. Then give the others a different primary: a more specific title and H1 is what moves the inference, and a link from the losing pages to the winner is what moves Google.
Read next
Site settings