SEO has moved from helping crawlers find and rank pages toward earning useful visibility across conventional results, generated summaries, and conversational interfaces. The practical response is not to abandon technical and editorial fundamentals. It is to measure each surface at the data grain it can actually support.
SEO, or search engine optimization, is the work of improving a site’s discoverability in unpaid search through accessible pages, useful content, and signals that help search systems understand and assess it. The history below separates documented product facts from disputed attribution and third-party estimates, then turns those distinctions into an operating workflow.
A Search Console impression is an aggregate search-performance measure. A recorded AI-answer citation is one observation. Neither, by itself, proves a visit, a conversion, or a sale.
How SEO evolved and what teams should measure now
Early search tools helped people locate files and information. As Web search developed, publishers learned to influence rankings with keywords, metadata, and links. Search engines responded with broader systems for assessing relevance, usefulness, and quality. Discovery now also happens through generated answers and multimodal interfaces. More surfaces make one ranking or traffic metric less sufficient, not traditional SEO obsolete.
A search metric is meaningful only when its source, population, and data grain are clear.
Before SEO: search tools and the early Web
Archie, created by Alan Emtage at McGill University, is widely regarded as one of the earliest Internet search tools. It indexed file names in anonymous FTP archives, not the contents of modern Web pages. Calling it “the first search engine” depends on what counts as a search engine and which network or information system is included. McGill’s institutional history of Archie supports the qualified description, not every possible “first” claim.
In 1945, Vannevar Bush described the Memex, a proposed device for storing and associating knowledge. It is useful context for ideas about retrieval and links, but it was not the Web’s direct cause. Bush’s essay in The Atlantic documents the proposal.
PageRank and the shift toward link quality
Google’s original system used link structure and the relative importance of linking pages as part of its approach to search. The Stanford paper by Sergey Brin and Larry Page describes PageRank within Google’s early search architecture. It supports saying that link analysis became central to Google, not that PageRank was the first algorithm ever to consider link quality. Read the Stanford paper on Google’s early search engine.
Links are signals in a broader system, not simple votes that guarantee a position. Their context and quality matter. A signal that can be counted can also be manipulated, which is why link relevance and editorial judgment remain more useful than raw volume.
From easily manipulated signals to quality and trust
When ranking systems relied heavily on repeated terms or link counts, publishers had incentives to stuff keywords, create doorway pages, and exchange links to inflate signals. Search evolved toward assessing content and links in broader context, alongside whether pages can be accessed and understood. SEO-industry names for updates do not always correspond to a complete official account of each change, so familiar labels should not be treated as precisely documented penalties without evidence.
Google introduced nofollow in January 2005 to help address comment spam. In 2019, Google described nofollow, sponsored, and ugc attributes as hints rather than absolute directives. The original nofollow announcement documents its introduction, while Google’s 2019 guidance describes the later treatment.
Mobile-first indexing and E-E-A-T in practice
Mobile-first indexing means Google primarily uses the mobile version of a site’s content for crawling and indexing. It is not a separate index. Google completed the transition in 2023. For a site with different mobile and desktop versions, check that important content available on desktop is also available to Google on mobile. Google’s mobile-first indexing update explains the status and practical implication.
E-E-A-T means Experience, Expertise, Authoritativeness, and Trustworthiness. Google says E-E-A-T itself is not a specific ranking factor. Quality raters assess search results, but their evaluations do not directly set rankings. Treat E-E-A-T as a quality framework, especially when readers need trustworthy information, rather than as a score or isolated ranking switch. Google’s helpful-content guidance explains the distinction.
AI search adds discovery surfaces, not a verified new ranking report
Google’s Search Generative Experience was an experiment. The company later introduced AI Overviews and announced AI Mode as another Search experience. In May 2026, Google announced multimodal Search inputs and generated interfaces. The announcement described staged availability, so an announced capability should not be assumed to be available to every user, region, language, account, or subscription.
Google’s AI Mode announcement provides the initial product context, while its May 2026 Search announcement describes later multimodal inputs and generated interfaces. Google’s separate source-discovery announcement explains how links can appear in AI Search experiences. These announcements establish product direction and stated capabilities, not a universal site-owner submission mechanism or public citation-ranking API.
The operating response is to maintain useful, crawlable, well-supported content and separately observe AI-search appearances where collection is permitted. A feature announcement belongs in a dated availability record, not an assumption that every reader can use the feature.
Clicks, impressions, and AI citations answer different questions
A bounded example shows why measurement needs context. In a Pew Research Center study of browsing data from 900 U.S. adults covering March 2025, users clicked a traditional result on 8% of visits when an AI summary appeared, compared with 15% when it did not. Links inside the summary were clicked on 1% of visits. These findings describe that sample and period, not a universal or current click-through rate. Read Pew’s study and methodology.
| Data source | One record represents | Cannot establish |
|---|---|---|
| Search Console | An aggregate for selected dimensions | Every individual search event |
| Prompt observation | One prompt run in a defined setting | Typical visibility across all runs |
| Citation log | One citation in one observed answer | A click, ranking, or conversion |
| Visibility summary | A calculation over a declared run population | Results outside that population |
Build a Search Console reporting workflow with explicit data grain
Google’s Search Analytics API returns aggregate clicks, impressions, CTR, and average position for an authorized property, subject to its documented dimensions and limitations. A row grouped by date is not equivalent to a row grouped by date, page, query, country, and device. Store the request definition with the result so later comparisons use the same grain.
The following is an illustrative request configuration, not a Google-prescribed storage schema or schedule. Authenticate for the property, call the Search Analytics query method, and retain the response metadata.
{
"property": "sc-domain:example.com",
"startDate": "2026-09-01",
"endDate": "2026-09-30",
"type": "web",
"dimensions": [
"date",
"page",
"query",
"country",
"device"
],
"rowLimit": 25000
}For each returned aggregate, preserve the property, search type, date range, dimensions, filters, aggregation type, dimension values, extraction time, and response metadata. A proposed uniqueness key is:
{
"source_system": "search_console",
"property": "sc-domain:example.com",
"search_type": "web",
"start_date": "2026-09-01",
"end_date": "2026-09-30",
"dimensions_hash": "declared-dimension-set",
"filters_hash": "declared-filter-set",
"aggregation_type": "auto",
"dimension_values": "all-returned-values"
}This is an implementation design, not a Google-defined key. If daily and query-level rows share a reporting store, include an explicit grain field and never identify a row only by page and date when query, country, device, or search appearance can vary.
Search Console data is delayed, aggregated, and quota-limited; the API may not return every row. Google documents the query method, data extraction guidance, API limits, and reported data anomalies. Check those constraints before interpreting a sudden change.
Observe AI-search visibility without pretending it is an official export
No general-purpose official Google export for AI Overview citations or AI Mode brand visibility was verified in the reviewed sources. A team can still design a governed observation log, subject to product terms, collection permissions, and privacy requirements. The design below is illustrative, not a vendor schema.
Keep three record types separate:
prompt_runmeans one execution of one prompt in one environment.citation_observationmeans one cited URL or entity within that run.visibility_summarymeans an aggregate over a declared prompt set, period, surface, and calculation.
Each run should record an ID, engine, surface, model variant if exposed, timestamp, country, language, exact prompt, collector, device or browser context where relevant, and an answer-capture reference. Each citation should retain its parent run ID, ordinal, normalized URL, displayed title or entity candidate, and review status.
{
"prompt_run": {
"run_id": "run-2026-10-13-001",
"engine": "Google",
"surface": "AI Mode",
"model_variant": null,
"prompt": "How do teams measure search visibility?",
"observed_at": "2026-10-13T14:00:00Z",
"country": "US",
"language": "en",
"collector_id": "analyst-01",
"answer_capture_ref": "restricted://observation/run-2026-10-13-001"
},
"citation_observation": {
"citation_id": "run-2026-10-13-001-01",
"run_id": "run-2026-10-13-001",
"citation_ordinal": 1,
"cited_url": "https://example.com/guide",
"entity_match": "candidate brand",
"review_status": "needs_review"
}
}One answer is one observation, not share of voice. A share-of-voice summary needs a declared prompt set, period, surface, locale, and calculation that another analyst can reproduce.
Repeated prompts can produce different answers, so absence in one run does not establish general absence. Use a database-enforced unique constraint and transactional upsert where supported. A read-then-insert check alone can race when concurrent jobs replay an observation. A practical citation key can combine the run ID, citation ordinal, and normalized cited URL; if the product can reorder or duplicate citations, add an event UUID and retain the raw capture for review.
Route ambiguous brand or entity matches to human review. If permitted, bounded AI assistance may propose candidate URLs or mentions from a captured answer, but it should not invent missing citations or resolve uncertain matches without review. For teams exploring that narrow support role, see AI agent implementation services.
Turn measurement into a decision, not a vanity dashboard
Use impressions to ask whether exposure changed, clicks and CTR to assess response, and conversion records in the analytics or business system to evaluate outcomes. Use prompt observations only to describe sampled AI-answer presence. Keep the systems of record explicit: Search Console for its search aggregates, the analytics or business platform for its own conversion data, and a separately governed store for prompt observations.
Assign an SEO measurement owner to investigate material changes and a named approver to decide whether to change content, technical access, or measurement coverage. Alerts should trigger investigation, not automatic CMS edits. A citation can inform an exposure hypothesis, but traffic or revenue attribution requires an independently defined method.
- Confirm the source system, record grain, and aggregation scope.
- Check the date range and, for rates, the numerator and denominator.
- Compare matching dimensions, filters, locales, surfaces, and collection conditions.
- Name the owner who investigates API errors, anomalies, or ambiguous observations.
- Agree who approves the next content, technical, or measurement action.
Frequently asked questions about SEO history
Who coined the term SEO?
The exact origin is disputed. John Audette later recalled proposing “Search Engine Optimization” in 1997. That is a firsthand recollection, not definitive documentary proof that one person invented the practice. Read Audette’s interview.
Is E-E-A-T a ranking factor?
Google says E-E-A-T itself is not a specific ranking factor. Use it as a framework for evaluating content quality and trustworthiness, not as a promised score.
Is SEO still useful?
Yes. Search is evolving, and SEO remains useful for making pages accessible, understandable, and valuable. Rankings and clicks still matter, but they do not describe every form of discovery.
Does the SEO 80/20 rule describe every site?
No. Organic traffic may be concentrated among a small share of pages or queries, but the pattern varies by site, market, and period. Treat 80/20 as a hypothesis to test against your own data, not a universal law.