Content and search intent: researching demand, choosing pages, writing them and keeping them current

Content SEO is the part of search work that decides what a site says: which questions it answers, which pages it has, how they are written and how they are kept true over time. It sits on top of the technical layer. A page has to be crawlable and indexable before any of this matters, and on most small sites the technical and local foundations are the real bottleneck, which is the argument of most local businesses do not need more content.

This guide follows the order the work happens in: research demand, read intent, decide which pages to create, brief them, write them, and then audit, update and measure. I tie claims about how search engines and assistants behave to their own documentation, claims about how people read and search to published studies with their sample and year, and mark my own judgement as judgement. Where industry data and a platform's statements disagree, both are shown. Products and figures change, so read every dated statement as true in October 2026.

Demand used to arrive as short queries typed into a search box. It now arrives in three shapes at once: short queries, full questions, and long prompts typed or spoken into an assistant or an AI search mode. The content work starts with knowing which shape your audience uses for which job, because the same need can be expressed as “emergency plumber leeds”, “how much does an emergency plumber cost at night” and “my boiler is leaking from the bottom and the pressure keeps dropping, what should I do before a plumber gets here”.

Google said at its 2025 developer conference that people were coming to Search with “more complex, longer and multimodal questions”, and claimed that in the US and India AI Overviews were driving over 10% more usage of Google for the types of queries that show them.¹ That is the platform's own measure of its own product, so treat it as a direction of travel. Independent data points the same way on query shape. Pew Research Center tracked the Google searches of 900 US adults in March 2025, 68,879 searches in all. An AI summary appeared on 18% of them, but on 60% of searches phrased as questions and 53% of searches of ten words or more, against 8% of one or two word searches.²

Assistants add a different kind of demand. An NBER working paper by OpenAI and academic economists, published in September 2025, classified a large sample of ChatGPT conversations and found that non-work messages had grown from 53% to more than 70% of usage, and that practical guidance, seeking information and writing together made up about 80% of conversations.³ When an assistant needs the web, it does not search with the prompt as typed. OpenAI documents ChatGPT search rewriting a prompt into one or more targeted queries and sometimes sending more specific follow-up queries after reading the first results.⁴ Your page is found by those rewritten queries, so the ordinary work of covering a topic clearly is what gets it retrieved.

The practical consequence is that prompt-shaped demand cannot be researched with a volume column. You find it in the words customers already use with you: enquiry forms, sales call notes, live chat, reviews and community threads. Each prompt still breaks down into needs a page can meet: a comparison, a price, a limit, a fit for a condition. Content that states those facts plainly serves all three shapes at once. The AI search pillar covers how assistants turn prompts into searches in detail, and the tutorial on getting recommended by AI assistants shows how to build a set of prompts to track.

Every keyword tool sells the same promise: a list of what people search and how often. The list is useful. The numbers beside it need more care than most reports give them, and the gaps in the list matter as much as what is in it.

Where the volume numbers come from

Google's Keyword Planner reports average monthly searches for a keyword and its close variants, and the figures are rounded, so volumes across locations may not add up.⁵ It was built to plan advertising, which is why it groups variants together. Third-party tools model their volumes from a mix of clickstream panels, Keyword Planner data and ranking data, and none of them sees Google's logs. Ahrefs compared Keyword Planner volumes with Search Console impressions for keywords where sites held top positions, and found Keyword Planner roughly accurate in 45% of cases. That study is from 2021 and was run by a company selling a competing tool, so read it as evidence that volumes are estimates, which no one disputes.⁶

Google Trends is a different instrument. It divides each data point by total searches in that place and time, scales the result from 0 to 100, and uses a sample of searches.⁷ It shows relative interest and seasonality well, and says nothing about absolute volume. Use it to see whether a topic is growing, when it peaks in the year, and how two phrasings compare.

The long tail is most of the demand, and most of it is invisible

Ahrefs reported in May 2026 that keywords with fewer than 10 searches a month make up almost 93% of its US keyword database of about 28.7 billion keywords.⁸ Ahrefs defines long-tail keywords by volume, not word count, which is the useful definition. Tools record those low-volume phrases as zero or leave them out. A page that ranks well for a topic picks up hundreds of them. In the same 2021 study, the top page for “submit website to search engines” ranked for about 1,400 keywords.⁶ The implication for planning is that you choose topics, not keywords, and judge a topic by the total demand around it.

What Search Console tells you that no tool can

Search Console's Performance report shows the queries your site actually appeared for, with impressions, clicks and average position. It has its own gaps: anonymised queries are omitted from the tables, though they are counted in the chart totals.⁹ It supports regular expression filters using RE2 syntax, which match anywhere in the string by default and ignore case unless you add (?-i).⁹ Two filters do most of the research work: one that isolates question queries, and one that separates brand from non-brand.

# Questions: queries that start with a question word
^(how|what|why|when|where|which|who|can|does|do|is|are|should|will)\b

# Long queries: eight words or more
^(\S+\s+){7,}\S+$

# Brand: use with "Doesn't match regex" to see non-brand demand
(acme|acme ?hire|acmehire)

The long-query filter is the nearest thing to prompt data that Search Console offers. Bing Webmaster Tools goes further: its AI Performance report, in public preview since February 2026, shows how often your pages are cited in AI answers and the grounding queries the system used when it retrieved them.¹⁰ Those grounding queries are the rewritten searches discussed above, and they are the best public evidence of how an assistant turns a prompt into a search.

Search intent is what the searcher wants to achieve. The usual labels are informational, commercial investigation, transactional and navigational, and a local intent cuts across all four. The labels help in a spreadsheet. The results page is better evidence, because it shows what the search engine has learned people want for that exact query, in that country, this month.

Ahrefs studied about 14 billion pages in its index in 2023 and found that 96.55% got no traffic from Google. Its three explanations were no search demand, no backlinks, and a mismatch with search intent, with the example of a product page that will not rank for “best yoga mats” because searchers want a comparison.¹¹ The sample leans towards pages Ahrefs crawls, so the exact share is less important than the third reason, which is the one content teams control.

What to look at on the page

• The format of the top results. Guides, lists, product categories, single products, tools, videos, forum threads or local listings. If eight of ten are category pages, an article will struggle.
• The angle. “Best”, “cheap”, “for beginners”, “2026”, “near me”. Titles that repeat one angle across the page tell you what the searcher is filtering on.
• The features. AI Overview, People Also Ask, a local pack, shopping results, video, images, Discussions and forums. Each changes how many clicks are left and which kind of page can earn them.
• Who ranks. Brands, publishers, marketplaces, forums or small specialists. A page full of forum threads means people want first-hand experience.
• The depth. Whether winning pages answer in a paragraph or cover the subject in a long guide. Copy the depth the query needs, not the word count of the winners.
• Mixed intent. Some results pages split between, say, guides and shops. That usually means the query is ambiguous, and one page cannot serve both halves well.

AI Overviews have moved beyond informational queries, which changes how you read intent from features. Semrush analysed over 10 million keywords across 2025 and found the share of keywords showing an AI Overview went from 6.49% in January to a peak of 24.61% in July and 15.69% in November. Of the keywords that triggered one, 91.3% were informational in January and 57.1% in October, with commercial, transactional and navigational queries making up the rest.¹² Semrush sells SEO software and the figures depend on its keyword set, but the direction matters: an AI Overview on a commercial query is now normal, and a page that wins that query has to give the facts an overview would draw on.

How many clicks are left

The studies disagree on how much an AI summary reduces clicks, and the disagreement is worth knowing. Pew found people clicked a traditional result in 8% of visits with an AI summary against 15% without one, clicked a link inside the summary in 1% of visits, and ended their browsing session on 26% of pages with a summary against 16% without.² Ahrefs compared 300,000 keywords using aggregated desktop Search Console data and found an AI Overview correlated with a 58% lower average click-through rate for the top-ranking page in December 2025.¹³ Semrush, using clickstream data from Datos, which Semrush owns, on a smaller set of keywords, compared the same terms before and after an AI Overview appeared and found the zero-click rate fell slightly, from 33.75% to 31.53%.¹² The methods differ, so the numbers cannot be lined up against each other. For planning, read a results page with an AI Overview, a local pack and People Also Ask as one where the remaining clicks go to pages that offer something the summary cannot: a tool, a price, a comparison, a decision.

A content plan is a list of decisions about pages. Each page should exist because a distinct group of people needs something the site can give, and no other page on the site already gives it. That rules out most of what a raw keyword list suggests. Keyword lists are full of variants that one page should cover, and of topics the business has no standing to write about.

A topic map

A topic map groups demand by the jobs customers are doing and assigns each group to one page. The top of the map is what the business sells or knows. Below it sit the decisions people make on the way to buying, and below those the questions that come up along the way. Each node is one page, with one primary intent, and the links between pages follow the path a reader takes. The SaaS and ecommerce pillars show how this looks for product sites; on a local site the map is usually small, and the location pages guide covers where it tips into doorways.

Tests for whether a page deserves to exist

• Distinct intent. The results page for this topic differs from the results page for a topic you already cover. If the same pages rank for both, it is one page.
• Something to add. You have first-hand experience, data, prices, examples or a view that the current results lack. If the best you can do is summarise what ranks, the page will not earn its place.
• A fit with the business. The topic is one your customers need on the way to buying, or one where your expertise is real. Traffic from unrelated topics rarely converts, and publishing on many unrelated topics in the hope that some perform is a listed warning sign of search-engine-first content.¹⁴
• Someone to maintain it. Prices, specifications, laws and product names change. A page nobody will update becomes a liability.
• A route to it. You can link to it from relevant pages, and you know how it will earn links or mentions if the topic is competitive.

Publishing pages for every variation of how people might phrase something is a specific risk now. Google's guide to generative AI search says that creating separate content for every possible variation, including fan-out queries, primarily to manipulate rankings violates its scaled content abuse policy.¹⁵ Bing's rewritten guidelines, reported in February 2026, renamed the keyword stuffing section “Keyword Stuffing and Artificially Engineered Language” to cover content designed to manipulate AI answers.¹⁶ One thorough page per intent is both safer and more likely to be retrieved.

Cannibalisation is the label for two or more pages on one site competing for the same query. The term is overused. Search Console shows pages that were actually shown, and seeing several of your URLs for one query is often fine. When asked about this on Bluesky in September 2025, Google's John Mueller said that three different pages appearing in the same results did not seem problematic to him just because it was more than one, and that you need to look at the details.¹⁷

It becomes a real problem when the pages share an intent and split the signals one page should have. The symptoms are recognisable: the ranking URL for a query flips between two pages from week to week, neither page ranks as well as the topic deserves, and the two pages say much the same thing. Google's deduplication systems show only the most relevant of very similar results, so near-duplicate pages compete for a single slot.¹⁸

A brief is where the research turns into a page. A good one makes the writer's job about knowledge and judgement, because the intent, the scope and the facts to include are settled before the first sentence. A bad one is a keyword list and a word count. There is no preferred word count, and writing to one is a listed warning sign of search-engine-first content.¹⁴ Length alone does not matter for ranking; there is no minimum or maximum.¹⁹

What goes in a brief

The most important field is “what we add”. The original content and reviews systems are designed to reward original reporting, analysis and research, and to show original content ahead of pages that only cite it.¹⁸ The guidance for AI features makes the same point: draw on what you know and the in-depth experience you can bring, and do not just recycle what others have said.¹⁵ If the brief cannot fill that field, the page probably should not be written yet.

People do not read web pages the way they read a book, and retrieval systems do not read them whole either. The good news is that the habits that help one help the other. The research on readers goes back almost thirty years and has held up.

How people read

In Nielsen Norman Group's 1997 study, 79% of test users always scanned a new page and only 16% read word by word. Rewriting a site to be concise improved measured usability by 58%, making it scannable by 47%, making the language objective by 27%, and doing all three by 124%.²⁰ Nielsen later estimated that users read at most 28% of the words on a page during an average visit, while noting that good information architecture, layout and writing can turn scanning into reading.²¹ Eyetracking showed the F-shaped scanning pattern in 2006, and a 2017 retest found it still common on desktop and mobile, along with a layer-cake pattern in which people read headings and skip the text between them.²² The writing advice that follows is plain: put the most important information first, start headings with the words that carry information, and make the first sentence of each paragraph the most important one.²²²³

How retrieval systems read

Search engines have scored parts of pages for years. Google's passage ranking is an AI system that identifies individual sections of a page to understand how relevant the page is to a search.¹⁸ Retrieval-augmented systems go further: they split documents into chunks, find the chunks most relevant to a query, and write from those. OpenAI's file search, a developer product, splits files into 800-token chunks with a 400-token overlap by default. Its documentation also describes rewriting the query before searching, weighing matches by meaning against matches on the exact words, and a score threshold below which chunks are not returned.²⁴ Anthropic's research on contextual retrieval shows the weakness of that approach: a chunk that says “the company's revenue grew by 3% over the previous quarter” loses the company and the quarter once it is cut from its document. Adding context to each chunk before indexing reduced failed retrievals by 35%, by 49% when combined with keyword matching, and by 67% with reranking added, in Anthropic's tests, where a failure was a relevant chunk missing from the top 20 results.²⁵

Neither of those is how Google Search or ChatGPT search works internally, and no provider publishes that. They show the general problem. A passage that only makes sense with the paragraph before it is harder to retrieve and easier to misquote. For Google's AI features there is no requirement to break content into tiny pieces, and no special files or markup are needed.¹⁵ What helps is the same thing that helps a scanning reader: sections that name their subject and state their facts.

What the GEO research found

The closest thing to a controlled test of writing for AI answers is the GEO paper (Aggarwal and others, KDD 2024). The authors rewrote source pages nine ways and measured how much of a generated answer drew on each. Adding quotations, statistics or citations to sources gave relative improvements of about 30 to 40% on their visibility measures, and keyword stuffing gave little or no improvement. Lower-ranked sources gained most: citing sources increased visibility for the fifth-ranked source by 115.1%.²⁶ The tests used GPT-3.5 with search results and Perplexity, on a benchmark the authors built, so read the paper as evidence for a direction, not a recipe for any current product. The direction matches Bing's own advice for AI citations: clear headings, tables and FAQ sections, fresh and accurate content, and examples, data and cited sources.¹⁰

Habits that serve both

• Answer first. Put the answer to the page's main question in the first two or three sentences, then explain.
• One idea per section, with a heading that says what the section answers. “Prices for an e-bike service” beats “Our approach”.
• Repeat the subject. Write “an e-bike service costs £95” in place of “it costs £95”, so the sentence survives being quoted alone.
• State facts specifically: prices, dimensions, limits, dates, places, conditions. Vague claims give a reader and a retrieval system nothing to use.
• Use tables for comparisons and lists for steps or options. They are easy to scan and easy to extract.
• Cite sources for claims that are not yours, and say where your own numbers come from.
• Cut what adds nothing. Introductions that restate the title and conclusions that restate the page both cost the reader attention.

E-E-A-T stands for experience, expertise, authoritativeness and trust. It comes from the guidelines Google gives its search quality raters, people who assess sample results so Google can measure how well its systems are working. The raters' overview describes the four parts as the first-hand experience of the creator, the expertise of the creator, the authoritativeness of the creator, the content and the website, and trust: the extent to which the page is accurate, honest, safe and reliable.²⁷ Pages on “Your Money or Your Life” topics, where poor information could harm someone's health, finances or safety, are held to a higher standard.²⁷

Two points from the same sources keep the idea in proportion. Rater scores do not directly change how any page ranks; they are aggregated to measure how well the algorithms are performing.²⁷ And asked whether E-E-A-T is a ranking factor, the SEO starter guide answers “No, it's not.”¹⁹ The helpful content guidance puts it more precisely: E-E-A-T itself is not a specific ranking factor, but Google's systems use a mix of factors to identify content that demonstrates it, give more weight to it on topics that could affect health, finances or safety, and treat trust as the most important part. The other three contribute to trust, and content does not have to show all of them.¹⁴

Showing first-hand experience honestly

First-hand experience is the part of E-E-A-T most pages can actually improve, and the part AI tools cannot supply. A first-hand review offers a unique perspective based on personal experience; a summary of existing content restates what is already available elsewhere.¹⁵ On the page that means things a reader can check: your own photos and measurements, what you tested and how, what went wrong, prices you paid, results from your own work, and a clear statement of who wrote it and why they know.

• A byline that names a real person, with a short author page saying what they do and why they know the subject. It should be self-evident who wrote the content, pages should carry a byline where one is expected, and bylines should lead to background on the author.¹⁴
• Original evidence: your photos, screenshots, test results, data, and the method behind them.
• Clear sourcing for claims that are not yours, one of the things that makes people want to trust content.¹⁴
• Specifics only a practitioner would include: the mistake people make, the part that fails first, the case where the usual advice does not apply.
• A review by a qualified person on health, finance, legal and safety topics, named on the page.

Using AI to help produce content is not against any search engine's rules. Google's guidance on AI-generated content, published in February 2023, says it rewards high-quality content however it is produced, and that using automation to manipulate rankings is spam.²⁸ The rule sits in the spam policies: scaled content abuse is when many pages are generated for the primary purpose of manipulating search rankings and not helping users, and the examples include using generative AI tools to generate many pages without adding value for users, and stitching or combining content from different web pages without adding value.²⁹

Bing moved in the same direction in its February 2026 rewrite. Its older guidelines said machine-generated content would result in penalties; the new wording, as reported, says large-scale content generated without oversight, quality control or editorial review often lacks usefulness, accuracy and originality, and may be excluded from indexing.¹⁶ Both engines are describing the same failure: volume without editorial judgement. Neither is describing a person using a model to draft, tidy or translate a page they then check and stand behind.

Google's helpful content guidance also asks about disclosure. Its “Who, How, Why” questions include whether the use of automation, including AI generation, is self-evident to visitors through disclosures, and whether you give background on how it was used.¹⁴ That is a self-assessment question, not a labelling requirement, but where a reader would reasonably expect to know, such as an automatically generated summary or a translated page, saying so is part of being trustworthy.

In the UK, fake reviews and concealed incentivised reviews became a banned practice under the Digital Markets, Competition and Consumers Act 2024, which the CMA describes as automatically unfair and illegal; its guidance also requires traders that publish reviews to take steps to prevent fake ones.³⁰ Other countries have their own rules, so take advice for your market. The translation case is covered in more depth in localising content for international SEO, which sets out a review tier for each kind of page.

A workflow that keeps AI drafting on the right side of the line

The content editing skill in the site's free skill pack for AI coding agents follows the same rules: it edits one real page for clarity and answerability, and will not invent statistics, quotes, sources or expertise, or mass-generate pages.

On-page elements get more attention than they deserve in many audits, and less care than they deserve on many pages. They will not make a weak page rank. They do decide how a page is presented in results, how easy it is to scan, and how well search engines understand how pages relate.

Title elements

Google builds the title link in results from several sources: the title element, the main visible title, headings, og:title, other prominent text, anchor text in links pointing at the page, and WebSite structured data.³¹ There is no limit on how long a title element can be, but the title link is truncated to fit the device.³¹ The best practices are short: give every page a title, make it descriptive and concise, avoid keyword stuffing and boilerplate repeated across pages, and brand it concisely.³¹ If Google rewrites your titles often, it is usually because they are long, stuffed, repeated across pages or do not match the visible heading.

The HTTP Archive's 2024 Web Almanac, which analyses its crawl of home pages plus one inner page per site, found title elements on 98% of desktop pages and 98.2% of mobile pages, with a median length of 12 words and 77 characters on desktop.³² A median title that long will often be cut in results, which is a reason to put the words that matter first.

<title>E-bike servicing in Leeds: prices and what is included | Spoke & Chain</title>
<meta name="description" content="Full e-bike service £95, with motor diagnostic and firmware update for Bosch and Shimano. Collection across Leeds. Book online or call.">
...
<h1>E-bike servicing in Leeds</h1>

Meta descriptions

Snippets are mostly built from page content, and Google sometimes uses the meta description when it gives a more accurate summary of the page.³³ There is no length limit, and the snippet is truncated as needed.³³ Write a unique description per important page that states the page's specific facts: price, scope, location, date. The Web Almanac found meta descriptions on 66.7% of desktop pages in 2024, so a third of pages leave the snippet entirely to the search engine.³²

Headings

Headings are for readers first: they are what scanners read, and what the layer-cake pattern relies on.²² For Google Search, headings used out of order do not matter.¹⁹ Use one clear main heading that matches the topic of the page, then section headings that say what each section answers. The Web Almanac found an h1 on 70% of pages, and an empty h1 on 6% of desktop pages, which is usually a template fault worth fixing.³²

Internal links

Internal links tell readers and crawlers which pages matter and how they relate. Google can only follow a link that is an a element with an href attribute, and descriptive anchor text helps people and Google understand the page linked to; “click here” does not.³⁴ On a content site the useful pattern follows the topic map: each page links up to the page for the bigger job, across to the pages for closely related questions, and down to the detail. A new page should get links from the pages that already rank for its parent topic on the day it is published.

• Link from relevant body text, with anchor text that says what the target is about.
• Make sure every page you want found has at least one link from a page that is crawled often.
• When you merge or redirect a page, update the links that pointed at it, so they go straight to the new URL.
• Do not stuff anchor text with keywords or repeat the same exact phrase across hundreds of links.
• Things that matter less than often claimed: keywords in the domain or URL path, which have hardly any effect beyond appearing in breadcrumbs, and the meta keywords tag, which Google Search does not use.¹⁹

A content audit gives every page one decision. It is the content equivalent of a stock take: most sites that have published for a few years carry pages that once had a purpose and no longer do, pages that compete with each other, and pages that are out of date. The audit finds them and decides what happens to each.

Gather the data

Decide one action per page

The mechanics of each action matter. A 404 or 410 tells Google the content does not exist, and a previously indexed URL is removed from the index; the two codes are handled the same way.³⁵ A 301 redirect is a strong signal that the target should be processed in place of the old URL, and a 302 is a weak one, so use 301s for merges.³⁵ Redirect only to a page that genuinely replaces the old one. A redirect to the home page or an unrelated page helps no one. The redirect mapping guide covers building the map, and the website migrations pillar covers doing it at scale.

The content audit skill in the skill pack runs this process with an AI coding agent: it inventories the site, assesses each page for quality, intent, cannibalisation and decay, and recommends one action per page. The positioning and strategy skill works on the step before, which topics a site should own.

Content decay is the slow loss of traffic from a page that once performed. The causes are ordinary: competitors publish better pages, the facts go out of date, the results page changes format, an AI Overview starts answering the question, or demand for the topic falls. Each needs a different response, so the first job is diagnosis.

Freshness matters for some queries more than others. Google runs “query deserves freshness” systems that show fresher content where it would be expected.¹⁸ A query about a product launch, a law change or this year's prices needs current content; a query about how a bicycle chain works does not. Updating a page that does not need it wastes effort. If a whole section of the site has dropped at once, start with the tutorial on recovering from a traffic drop before touching content.

Dating pages honestly

Changing the date on a page to make it look fresh when the content has not substantially changed is a listed warning sign of search-engine-first content.¹⁴ Good dating practice is a visible date labelled as published or last updated, structured data with datePublished and dateModified, and dates that describe when the page itself was published or updated.³⁶ Change the updated date when the substance changes: new facts, corrected advice, a rewritten section. Do not change it for a typo or a new image.

<script type="application/ld+json">
{
  "@context": "https://schema.org",
  "@type": "Article",
  "headline": "Why bike gears skip under load, and how to fix it",
  "datePublished": "2025-03-14T09:00:00+00:00",
  "dateModified": "2026-09-30T10:30:00+01:00",
  "author": {
    "@type": "Person",
    "name": "Sam Example",
    "url": "https://www.example.com/about/sam-example"
  }
}
</script>

Measure each page against the job it was given in the topic map. A guide written to capture early research is judged on non-brand impressions, clicks and the visits it sends to commercial pages. A service page is judged on enquiries. A comparison page is judged on sign-ups or sales it assists. Reporting every page on total traffic hides most of this, and it is the reason many content programmes look either better or worse than they are.

The core reports

• Search Console Performance report: impressions, clicks, click-through rate and average position, by page and query. Use the regex filters above to separate brand from non-brand and to watch question queries.⁹
• Search Console Generative AI performance report: impressions for links to your site in AI Overviews and AI Mode, by page, country, date and device. It does not show queries.³⁷
• Bing Webmaster Tools AI Performance: total citations, average cited pages per day, grounding queries and citations per page for Copilot and Bing's AI answers.¹⁰
• Analytics: organic sessions, engagement and key events per landing page. In GA4 an engaged session lasts longer than 10 seconds, has a key event, or has two or more page views.³⁸
• Your CRM or sales data: which pages appear in the paths of leads and customers. See measuring SaaS SEO in pipeline for the B2B version.

Read AI impressions as visibility, not traffic. The click studies above found fewer clicks to websites when a summary appears, with the exception of Semrush's before-and-after comparison, and none of them can tell you about your pages specifically.²¹³¹² Track whether your pages are cited, and for which topics, and treat clicks from those citations as a bonus. For prompts that never reach a search console, the guide to measuring visibility in AI search sets out a manual tracking method.

Reading changes fairly

• Compare like with like: the same months a year earlier for seasonal topics, and the same set of pages over time.
• Separate brand from non-brand. Brand demand moves with advertising and press, and content rarely drives it directly.
• Annotate changes: publish dates, updates, merges, redirects and known algorithm updates, so a change in the line has a likely cause beside it.
• Look at query spread as well as volume. A page that appears for more distinct queries over time is covering its topic better, even if clicks are flat because of AI Overviews.
• Do not credit a single update with a change until it holds for several weeks and other pages did not move the same way.