The AI SEO Workflow: Five Phases, Honest Costs, and What Quietly Breaks

AI SEO Workflow - 1
AI SEO Workflow - 2
AI SEO Workflow - 3

By Tom Morgan · Published and fact-checked 1 November 2026 · About 12 minutes to read

Key takeaways

  • Google’s own guide to AI search, published on 15 May 2026, says that optimizing for AI Overviews and AI Mode is still SEO. It also lists several popular tactics as unnecessary for Google.
  • Search Console now has a dedicated Generative AI report. It shows impressions only, not clicks, so it tells you where you appear, not what it earned you.
  • When an AI Overview is present, Ahrefs measured a 58% lower click-through rate for the top organic result (December 2025 data). Rankings can hold while traffic leaves.
  • Most of the workflow below costs a few hundred dollars a year if you work solo. The expensive part is the judgment you add, not the software.

Why most AI SEO advice is already out of date

Two things happened in 2026 that make older playbooks unreliable. Google published its first official guidance on generative AI features, and Search Console gained reporting for them. Before that, the field ran on guesses, vendor dashboards and confident tweets.

The guesses were not harmless. If you read AI SEO content from a year ago, you were probably told to publish an llms.txt file, split pages into “AI-friendly chunks,” add special markup, or rewrite everything for machines. Google’s guide, titled “Optimizing your website for generative AI features on Google Search,” says none of that is required for its features. Search Engine Journal’s summary covers the “mythbusting” section. The guide’s central line is that generative AI search optimization is the same job as search optimization, which is to say it’s still SEO.

That sounds reassuring, and it is, partly. The uncomfortable half is that the guide leaves your measurement problem untouched. Clicks are falling on pages that still rank, and a rank tracker will not show it.

This article lays out a five-phase workflow that respects both facts. Every number below is tied to a source I could check on 1 November 2026, and where the evidence is thin I say so.

What the data says right now

How often do AI Overviews appear? Nobody outside Google knows the real figure, and the estimates depend heavily on who is counting.

The reliable takeaway is the direction, and the strongest signal is in commercial queries: “best X for Y,” reviews, alternatives and comparisons. If you publish that kind of content, AI Overviews are already the norm in your niche, not an edge case.

What do they do to clicks? Ahrefs re-ran its 300,000-keyword study using December 2025 data and found that an AI Overview correlated with a 58% lower click-through rate for the page in position one. An earlier run in 2025 had found about 34.5%. The effect shrinks further down the page but doesn’t disappear. Separately, Pew Research Center’s browsing study (2025) found that people clicked a traditional result on 8% of visits when an AI summary appeared, versus 15% when it didn’t.

These are correlations from different methods, so don’t merge them into one number. They agree on the direction and on the rough size, and that’s enough to plan around.

The workflow at a glance

PhaseGoalMain outputTypical time
1. Entity and intent mappingKnow what you must cover and why searchers comeEntity map, intent labelsWeek 1
2. Brief and prioritizationDecide what to write firstRanked briefsWeek 2
3. Technical audit triageFix what costs you traffic, in orderPrioritized fix listWeek 3
4. Earning citationsBe the source an answer draws onOriginal assets, entity consistencyWeek 4
5. MeasurementSee what changed, honestlyAI-aware reportingOngoing

The week labels are a suggested pace for one small site, not a law. A large site will take longer.

Phase 1: Map entities and intent before you open a keyword tool

A keyword list tells you what people type. An entity map tells you what a credible source on the topic is expected to cover: the concepts, the named things, the processes, and the questions that follow them.

Start with three to five core entities for your site. An entity is a thing, not a phrase: “Core Web Vitals,” “freelance contract,” “prompt injection.” Then use a content-planning tool (MarketMuse and Clearscope both do this; see the cost table) or a manual SERP review to list the related subtopics and questions around each one.

Next, classify intent. A language model is a fast first pass for this, provided you check its work against the live results page:

text

Classify each entity below by dominant search intent:
informational, commercial investigation, transactional, or navigational.

For each one, give me:
1. The specific problem the searcher is trying to solve.
2. The question they are most likely to ask next.
3. The page format that tends to satisfy that intent.

Flag any entity where you are unsure, and say what you would
need to see in the live search results to decide.

The last instruction matters. A model will happily give a confident label for an ambiguous query, so ask it to show its uncertainty, then open the actual results and compare. If the first page is all product pages, an informational essay will not break in.

Where it breaks. Intent mismatch doesn’t announce itself. The page gets indexed and collects impressions. CTR stays low, and the usual reaction is to polish the page, when the real problem is that you built a buying guide for someone who is still learning. Low CTR looks identical whether the cause is rank, snippet or intent, so check intent first.

Phase 2: Turn the map into briefs, and rank them by momentum

The brief matters more than the draft. A weak brief produces a polished article for a question nobody is asking. A strong one produces a rough draft that answers the right question, and rough drafts are cheap to fix.

Before you commit hours to a topic, check whether interest in it is rising, flat or fading. Google Trends is the free source, and there are two things to know about it in 2026:

Trends reports relative interest on a 0 to 100 scale, not search volume. Use it to compare momentum, never to estimate traffic. Here is a small function that works on an exported CSV:

python

import numpy as np
import pandas as pd
from scipy.stats import linregress


def trend_momentum(series: pd.Series, min_points: int = 24) -> dict:
    """Fit a straight line to a Google Trends series (relative 0-100 values).

    Returns the slope per period and R-squared. A planning signal only:
    it says nothing about search volume or ranking potential.
    """
    clean = series.dropna().astype(float)
    if len(clean) < min_points:
        raise ValueError(f"Need at least {min_points} data points, got {len(clean)}")

    fit = linregress(np.arange(len(clean)), clean.to_numpy())
    return {
        "slope_per_period": float(fit.slope),
        "r_squared": float(fit.rvalue ** 2),
        "latest_vs_mean": float(clean.iloc[-1] / clean.mean()),
    }

A rising slope with a decent R-squared says the topic is gaining attention steadily. A high R-squared on a flat line says it’s stable, which is fine for evergreen work. Don’t treat any fixed threshold as a rule; calibrate it on topics in your own niche where you already know the outcome.

Content-scoring tools like Surfer and Clearscope belong here, at the drafting stage. They compare your draft with pages that already rank and suggest terms and structure. Treat the score as a checklist of what similar pages cover. It’s correlational, so a score of 90 is not a promise.

Phase 3: Triage the technical audit by traffic, not by volume of warnings

A crawler will hand you thousands of issues. The skill is sorting them by what costs you visits. Screaming Frog can export crawl data, and Search Console exports give you clicks and impressions per URL. Join them and rank by impact.

Two checks before you automate anything:

  1. There is no documented crawl cutoff at a certain load time. A claim that circulates in AI SEO content says pages slower than about 1.2 seconds are excluded from AI indexing. I could find no such rule in Google’s documentation. Page experience metrics like Core Web Vitals are signals within a broader system, and not a gate. The published “good” threshold for Largest Contentful Paint is 2.5 seconds, per web.dev.
  2. Lab numbers are not field numbers. LCP from a crawler’s PageSpeed integration is a lab measurement. Google’s assessment uses real-user field data, so check the Core Web Vitals report in Search Console before you spend a week fixing a lab-only problem.

Here is a prioritization function that doesn’t crash when a column is missing:

python

import numpy as np
import pandas as pd

DEFAULTS = {
    "status_code": 200,
    "lcp_ms": 0,
    "schema_valid": True,
    "crawl_depth": 5,
    "clicks": 0,
    "impressions": 0,
}


def prioritize_technical_issues(crawl: pd.DataFrame, gsc: pd.DataFrame) -> pd.DataFrame:
    """Merge a crawl export with Search Console data and rank issues by likely impact.

    Severity weights are a starting heuristic. Adjust them for your site.
    """
    df = crawl.merge(gsc, on="url", how="left")

    for col, default in DEFAULTS.items():
        if col not in df.columns:
            df[col] = default
        df[col] = df[col].fillna(default)

    df["crawl_depth"] = pd.to_numeric(df["crawl_depth"], errors="coerce").fillna(5).clip(lower=1)

    # Broken pages hurt most; redirects are usually fine unless chained.
    severity = np.where(df["status_code"] >= 400, 3, 0)
    severity = severity + np.where(df["status_code"].isin([301, 302, 307, 308]), 1, 0)
    severity = severity + np.where(df["lcp_ms"] > 2500, 2, 0)
    severity = severity + np.where(df["schema_valid"].eq(False), 1, 0)
    df["severity"] = severity

    traffic = df["clicks"] + 0.1 * df["impressions"]
    df["priority"] = df["severity"] * (1 + traffic) / df["crawl_depth"]

    return df[df["severity"] > 0].sort_values("priority", ascending=False)

On structured data. Valid markup still matters for the rich results Google supports, and errors are worth fixing. But Google’s AI guide says no special markup is needed for its generative features, so don’t oversell it to a client as an “AI Overview lever.” FAQ and HowTo rich results were also sharply curtailed in 2023, so check what Google currently supports before promising a result.

Phase 4: Earn citations by being worth citing

This is the phase where the industry says the most and knows the least, so it’s worth being precise about what is established.

Established. Google says its AI features run on its core ranking and quality systems. A page generally has to be indexed and eligible for a snippet to show up in them, and your property must not have opted out in Search Console’s generative AI setting. If a page can’t win a normal query, a summary is unlikely to cite it. So the foundation is the same foundation: crawlable, useful, original pages.

Plausible, not proven. For other assistants (ChatGPT, Perplexity, Claude and similar), there’s no equivalent official guide, and the research is young. A reasonable working hypothesis is that these systems favor sources they encounter repeatedly, specifically and consistently across independent sites, rather than a single impressive page. I’d call this a hypothesis to test, not a finding. If you try it, log the results.

What follows from both is practical:

  • Publish what a model can’t generate. Original data, first-hand tests, named case studies, screenshots of your own results, a defensible opinion. A page that restates the top five results gives a summarizer no reason to prefer it.
  • Keep your facts consistent everywhere. Product names, pricing, author names and claims should match across your site, your profiles and your press mentions. Contradictions are how you lose a citation you earned.
  • Earn mentions on pages that already rank. Digital PR, expert commentary and contributed data still work, and for the same old reason.
  • Skip the tricks Google has called unnecessary. That means llms.txt for Google, artificial chunking, and rewriting good pages in a “machine voice.”

One caution about measuring this. Citation visibility moves with model updates and with what competitors publish. A citation you hold today isn’t an asset on your balance sheet. Re-check your priority queries monthly.

Phase 5: Measure what changed, not what’s easy to count

This is where most teams get hurt, because the old dashboard keeps looking fine.

Use the new reports. Google rolled out its Generative AI performance reports to all sites worldwide on 31 August 2026. They show impressions from AI Overviews, AI Mode and generative features in Discover, by page, country and date. Two limits to state clearly in any report you send:

  • The report shows impressions, not clicks. You can see that you appeared. You can’t read off what that earned you.
  • In the standard Performance report, AI Overview and AI Mode traffic is counted inside the Web search type. There’s no filter that separates it. If you’ve read that the “Search type” filter isolates AI Overview traffic, that’s wrong.

Bing offers a comparable free view. Its AI Performance report in Webmaster Tools entered public preview in February 2026 and centers on citation frequency.

Add a canary. The signal worth watching is a query whose position holds while its click-through rate falls. That pattern points to something changing above you on the page, and an AI Overview appearing is one common cause. This function flags those queries from two Search Console exports:

python

import pandas as pd

REQUIRED = {"query", "impressions", "ctr", "prev_ctr", "position", "prev_position"}


def flag_ctr_drops(
    gsc: pd.DataFrame,
    min_impressions: int = 500,
    ctr_drop: float = 0.15,
    max_rank_move: float = 2.0,
) -> pd.DataFrame:
    """Flag queries where rank held but CTR fell.

    The output is a list of hypotheses to check against the live results page
    and the Generative AI report. It does not prove an AI Overview caused the drop.
    """
    missing = REQUIRED - set(gsc.columns)
    if missing:
        raise ValueError(f"Missing columns: {sorted(missing)}")

    df = gsc[(gsc["impressions"] >= min_impressions) & (gsc["prev_ctr"] > 0)].copy()
    df["ctr_change"] = (df["ctr"] - df["prev_ctr"]) / df["prev_ctr"]
    rank_held = (df["position"] - df["prev_position"]).abs() < max_rank_move

    return df[rank_held & (df["ctr_change"] <= -ctr_drop)].sort_values("ctr_change")

The 500-impression floor and the 15% drop are starting points. Seasonal queries will trigger false positives, so compare the same period year over year where you can.

Report conversions beside clicks. If AI Overview impressions are rising while pipeline is falling, that’s a conversion problem or a visibility-without-clicks problem. Those need different fixes, and a combined “SEO” number hides which one you have.

What it costs

I verified one price directly against the vendor’s own page. The rest come from third-party 2026 comparisons that disagree with each other, so I’m giving ranges. Plans change often, and a few of these tools have repriced several times in the last year. Check the vendor’s page before you budget.

ToolUse in this workflowApproximate entry costMain limitation
ChatGPT, Claude or similar assistantIntent classification, brief drafting, code helpAbout $20/month on common paid tiersConfident output with no built-in fact-check
Google Search ConsoleGenerative AI report, CTR monitoringFreeImpressions only for AI features
Bing Webmaster ToolsAI Performance (citation) reportFreePublic preview; narrower audience
Screaming Frog SEO SpiderCrawl audits£199 per year per user (about $279; vendor pricing); free up to 500 URLsDesktop software with a learning curve
SurferDraft scoring, on-page guidanceRoughly $79–$99/month at entry level, depending on source and billingScores correlate with ranking; they don’t cause it
ClearscopeTerm grading, editorial workflowRoughly $129–$189/month at entry levelPricey for solo operators; overlaps with Surfer
MarketMuseTopic modeling, content planningLimited free tier; paid entry roughly $99–$149/monthMethodology is not fully disclosed

A realistic solo stack: an assistant (about $20), one of Surfer or Clearscope, and Screaming Frog (about $23 a month spread across the year). That lands somewhere around $120–$230 a month depending on which scoring tool you choose. You don’t need both scoring tools. Their overlap is large enough that most solo operators should pick one.

I’m deliberately not including a traffic-value figure. Claims like “this workflow adds 8,000 to 15,000 visits a month” circulate widely, and none I’ve found names a client, a method and a timeframe. Your baseline, niche and competition will decide your result.

Four ways this fails without telling you

1. Great dashboard, dead funnel. Rankings hold, impressions hold, and top-of-funnel traffic slides because answers now appear above the click. The fix is to track CTR and conversions next to rank, and to review the canary list above monthly.

2. Using AI as the writer instead of the researcher. Google’s Search Essentials apply to AI-assisted content like any other, and its quality guidance targets pages that add nothing beyond what already exists. Use models for research, structure, code and first-pass analysis. Put your experience, data and opinion in the final draft.

3. Chasing tactics Google has already dismissed. If an agency’s AI SEO proposal centers on llms.txt, chunking or special markup for Google, ask which part of Google’s documentation supports it. Some of it may help with other systems. None of it is required for Google, and the cost of a bad bet is real hours.

4. Over-trusting one vendor’s tracker. AI visibility tools sample a limited set of prompts, and answers vary by session, location and model version. Treat their scores as directional and confirm anything important by hand.

Who should do what first

In-house SEO teams

Your stakeholders measure traffic and rankings. Both will look worse on informational queries even if your work is excellent, because the results page changed. Reset the conversation before the numbers move.

  1. Add the Generative AI report to your monthly review and record a baseline now.
  2. Report branded search, conversions and assisted revenue next to organic clicks.
  3. Build the canary report from Phase 5, and expect to spend real analyst time on it. Combining Search Console, analytics and CRM data into one story rarely works out of the box.
  4. Stop presenting AI Overview appearances as wins on their own. Tie them to what happens after the impression.

Freelance and independent strategists

Your leverage is judgment, and your tools are inexpensive.

  1. Sell the entity map and the one-page recommendation, not the 40-page export. The map shows effort, and the recommendation produces results.
  2. Run the solo stack above, and don’t buy both scoring tools.
  3. Put “what we’ll stop doing” in every proposal. Clients pay for fewer wasted articles.
  4. If a client’s method can be copied by buying the same subscriptions, your service isn’t differentiated yet.

What we still don’t know

Everything above assumes AI features on Google keep roughly their current shape. Coverage has expanded quickly, and Google has also pulled features back in sensitive areas (it restricted AI Overviews on some health searches in January 2026). Nobody can say whether coverage will plateau or keep climbing, and the two outcomes call for different budgets.

The sensible hedge is to keep investing in the fundamentals that feed both organic results and AI answers, measure honestly, and avoid bets that only pay off in one scenario. That is less exciting than a single bold prediction. It also survives being wrong.

Frequently asked questions

Is AI SEO different from regular SEO? For Google’s own features, Google says no: optimizing for generative AI search is part of SEO. The practical differences are in measurement and in the weight you put on original content, since summaries reduce the value of pages that only repeat what’s known.

Do I need an llms.txt file? Google’s guide says you don’t need new machine-readable files or AI-specific markup to appear in its generative features. Other systems may behave differently, but I haven’t found strong evidence that the file moves results for them either.

Can Search Console show my AI Overview clicks? Not separately. The Generative AI report shows impressions, and click data for AI features sits inside the Web search type without a dedicated filter.

How often do AI Overviews show up? It depends on the query set. Semrush found them on 88% of 400 tracked US commercial keywords in August 2026, while broader estimates from Similarweb and Semrush Sensor sit around 43% to 48%. Check your own queries in the Generative AI report and on the live results.

Which single tool should a solo operator buy first? Screaming Frog’s free tier plus one assistant gets you a working audit and intent workflow. Add one content-scoring tool when you’re publishing regularly.

How this article was sourced

The AI Overview figures come from Semrush (21 October 2026), Similarweb and Semrush Sensor as reported by Digital Applied, Ahrefs (December 2025 data) and Pew Research Center. Guidance on Google’s features comes from its Search Central documentation and Search Engine Journal’s coverage of it. Search Console rollout dates come from Search Engine Journal. Screaming Frog pricing comes from the vendor’s page; other tool prices are ranges drawn from third-party 2026 comparisons that conflict, and are labeled as such. Code samples were written for this article and require adapting to your column names. Nothing here is sponsored, and there are no affiliate links.

Tom Morgan runs content production across a group of niche publications covering AI tools, SEO, freelancing and content strategy. Corrections are welcome and will be dated here.

Leave a Reply

Your email address will not be published. Required fields are marked *